ggml-cpu: gelu optimizqation using simd

[ameya-ms]

Implementation:
Copies the same implementation as ggml_vec_silu_f32. Adding for AVX2, AVX512, NEON, and SVE. SSE2 uses the f16 table look-up.
Rewrote GELU function to only use exp instead of tanh.

Testing:

Edge cases vs scalar reference (0, ±inf, nan, saturated tails, tiny values,
and the scalar tail loop) pass on AVX-512, AVX2 and SSE2 builds.
Compiled and run on x86 AVX-512 / AVX2 / SSE4.2 builds. NEON/SVE compile-only
(not run — no ARM hardware).

SSE2 uses old version (everything else uses new version):
Numbers below are strictly comparing old/new version.

=== compiling gelu_sse2.exe ===
gelu_bench.cpp
=== compiling gelu_avx2.exe ===
gelu_bench.cpp
=== compiling gelu_avx512.exe ===
gelu_bench.cpp

#################### SSE2 ####################
N = 16777216 elements, 30 reps
table path: hardware F16C + precomputed f16->f32 table (matches ggml)

impl                   ms/run        Melem/s     max|err|
scalar tanh           265.859           63.1   0 (ref)
f16 table (old)        17.714          947.1   2.02e-03
SSE2 (new)             21.289          788.1   4.77e-07

speedup new vs table = 0.83x,  new vs scalar = 12.49x
checksum (ignore): 33037958.8

edge cases (SSE2 lane path vs scalar reference):
             x             simd           scalar   status
             0                0                0       ok
            -0               -0               -0       ok
             1         0.841192         0.841192       ok
            -1        -0.158808        -0.158808       ok
             5                5                5       ok
            -5     -2.29179e-07     -2.98023e-07       ok
            10               10               10       ok
           -10     -1.20409e-37               -0       ok
            20               20               20       ok
           -20               -0               -0       ok
           100              100              100       ok
          -100               -0               -0       ok
         1e-06            5e-07            5e-07       ok
        -1e-06           -5e-07           -5e-07       ok
           inf              inf              inf       ok
          -inf        -nan(ind)        -nan(ind)       ok
           nan              nan              nan       ok
tail test (n=1007, last 3 elems via scalar tail): max|err| = 4.77e-07   ok
edge-case result: ALL PASS

#################### AVX2 ####################
N = 16777216 elements, 30 reps
table path: hardware F16C + precomputed f16->f32 table (matches ggml)

impl                   ms/run        Melem/s     max|err|
scalar tanh           261.145           64.2   0 (ref)
f16 table (old)        17.836          940.7   2.02e-03
AVX2 (new)              8.384         2001.2   4.77e-07

speedup new vs table = 2.13x,  new vs scalar = 31.15x
checksum (ignore): 33037958.8

edge cases (AVX2 lane path vs scalar reference):
             x             simd           scalar   status
             0                0                0       ok
            -0               -0               -0       ok
             1         0.841192         0.841192       ok
            -1        -0.158808        -0.158808       ok
             5                5                5       ok
            -5      -2.2918e-07     -2.98023e-07       ok
            10               10               10       ok
           -10     -1.20409e-37               -0       ok
            20               20               20       ok
           -20               -0               -0       ok
           100              100              100       ok
          -100               -0               -0       ok
         1e-06            5e-07            5e-07       ok
        -1e-06           -5e-07           -5e-07       ok
           inf              inf              inf       ok
          -inf        -nan(ind)        -nan(ind)       ok
           nan              nan              nan       ok
tail test (n=1007, last 7 elems via scalar tail): max|err| = 4.77e-07   ok
edge-case result: ALL PASS

#################### AVX-512 ####################
N = 16777216 elements, 30 reps
table path: hardware F16C + precomputed f16->f32 table (matches ggml)

impl                   ms/run        Melem/s     max|err|
scalar tanh           264.979           63.3   0 (ref)
f16 table (old)        18.230          920.3   2.02e-03
AVX-512 (new)           6.771         2477.8   4.77e-07

speedup new vs table = 2.69x,  new vs scalar = 39.13x
checksum (ignore): 33037958.8

edge cases (AVX-512 lane path vs scalar reference):
             x             simd           scalar   status
             0                0                0       ok
            -0               -0               -0       ok
             1         0.841192         0.841192       ok
            -1        -0.158808        -0.158808       ok
             5                5                5       ok
            -5      -2.2918e-07     -2.98023e-07       ok
            10               10               10       ok
           -10     -1.20409e-37               -0       ok
            20               20               20       ok
           -20               -0               -0       ok
           100              100              100       ok
          -100               -0               -0       ok
         1e-06            5e-07            5e-07       ok
        -1e-06           -5e-07           -5e-07       ok
           inf              inf              inf       ok
          -inf        -nan(ind)        -nan(ind)       ok
           nan              nan              nan       ok
tail test (n=1007, last 15 elems via scalar tail): max|err| = 4.77e-07   ok
edge-case result: ALL PASS

Notice -inf -> -nan(ind)