feat: add functional per-head FP8 quantization for FA3

[happierpig]

This PR adds FP8 support in FA3 to speed up compute-bound prefill kernels. It follows up on #869.

1. Bug fixes

• Fixed deadlock, illegal memory access, and wrong results for varied num_heads and seq_len.
• Covered by unit tests.

2. New features

• Enabled FP8 in-kernel transpose logic of mainloop_sparse.cuh.
• FP8 now works in:
  • BatchPrefillWithPagedKVCache
  • BlockSparseAttentionWrapper: support sparse and quantized attention

3. Python JIT interface

• Exposed kernels to Python:
  • BlockSparseAttentionWrapper
  • single_prefill_with_kv_cache
• Migrated tests and benchmarks to Python scripts:
  • tests/test_hopper_fp8_attention.py
  • benchmarks/bench_hopper_fp8_attention.py

Note: Performance is on par with #869. Need tuning.

cc @yzh119

[yzh119]

Thank you @happierpig , great contribution!