Summary:

X-link: facebookresearch/FBGEMM#481

To have auto-vectorized code be competitive with asmjit we need to specialize the generic code to a some fixed parameters. We cannot specialize at runtime, so this introduce a framework to specialize for a given set of parameters at compile time and choose between existing specializations at runtime.

The framework added here allows to specify lines like the following for a given function.
Each parameter can be set to `var` to not specialize it or `fixed(C)` to create a specialized version with that parameter set to the constant value `C`. Example:

```
SPECIALIZE(
      /*BIT_RATE=*/fixed(2),
      /*BLOCK_SIZE=*/var,
      /*HAS_WEIGHT=*/fixed(true),
      /*NORMALIZE_BY_LENGTHS=*/var,
      /*PREFETCH=*/var,
      /*IS_WEIGHT_POSITIONAL=*/var,
      /*USE_OFFSETS=*/var,
      /*OUTPUT_STRIDE=*/fixed(int64_t{-1}),
      /*INPUT_STRIDE=*/fixed(int64_t{-1}),
      /*SCALE_BIAS_LAST=*/fixed(true),
      /*NO_BAG=*/fixed(false),
      /*IS_BF16_OUT=*/var,
      /*IS_BF16_IN=*/var)
```

This diff introduces some exemplary specialization for `GenerateEmbeddingSpMDMWithStrides_autovec` and `GenerateEmbeddingSpMDMNBitWithStrides_autovec` specializing them for bit_rate 2, 4 and block sizes 32, 64, 128.

This framework should make it easy to tune for common use-cases in production by specializing the commonly used parameters or remove specializations to conserve code size.

Reviewed By: excelle08

Differential Revision: D62984408