[float8nocompile] Make Float8Linear fp8 conversion kernel used configurable and update benchmarking code to compare both kernels

[danielvegamyhre]

Update Float8Linear and benchmarking code to easily compare the atomic_max-based fp8 conversion kernel vs the reduction based approach.

Example output:

input_shape    kernel_algo                 high_precision_dtype      eager_time    compiled_time    float8nocompile
-------------  --------------------------  ----------------------  ------------  ---------------  -----------------
(16, 4096)     KernelAlgorithm.ATOMIC_MAX  torch.bfloat16               649.288          409.496            409.756
(256, 4096)    KernelAlgorithm.ATOMIC_MAX  torch.bfloat16               684.32           421.179            444.348
(4096, 4096)   KernelAlgorithm.ATOMIC_MAX  torch.bfloat16              1827.66          1066.81            1178.63
(65536, 4096)  KernelAlgorithm.ATOMIC_MAX  torch.bfloat16             22343.4          12675.7            14644.2
(16, 4096)     KernelAlgorithm.REDUCTION   torch.bfloat16               648.859          408.961            562.557
(256, 4096)    KernelAlgorithm.REDUCTION   torch.bfloat16               684.449          421.08             612.861
(4096, 4096)   KernelAlgorithm.REDUCTION   torch.bfloat16              1827.6           1068.37            1604.26
(65536, 4096)  KernelAlgorithm.REDUCTION   torch.bfloat16             22434            12669.1            20523

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/1459

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 14d6a4f with merge base 567cb46 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.