TileLang DeepSeek-V4 porting

[MirkoDeVita98]

Adds PTO DSL ports of the six custom kernels used by DeepSeek-V4, plus
benchmarks and documentation. Each kernel follows the standard examples-tree
layout (compile.sh + run_*.py) and is exercised by
examples/validate_all_examples.py.

What's new — examples/aot/deepseek_v4/

Kernel	Pipe	What it does
act_quant/	vector	per-row absmax fp16 → int8 quant
fp4_act_quant/	vector	per-row fp16 → mxfp4 (e2m1) quant
fp8_gemm/	cube + vec	per-channel fp8 GEMM
fp4_gemm/	cube + vec	per-channel fp4 GEMM
hc_split_sinkhorn/	vector	fused MoE-router head: pre/post sigmoid + 20-iter Sinkhorn
sparse_attn/	vector	FlashAttention with indexed top-k KV gather + per-head sink logit

Each folder ships a <name>_builder.py, <name>_util.py, caller.cpp,
compile.sh, run_<name>.py, README.md, and .gitignore.

Notable design choices

• fp8_gemm / fp4_gemm keep the matmul pure (cube fp32 → fp16) and
  fold the per-channel Sa rescale into a host-side pre-scale of A,
  leaving only Sb on the vector pipe. Avoids two extra cube fragments
  per tile; matches reference within 5 × 10⁻³.
• hc_split_sinkhorn runs all three heads (pre / post / 20-iter
  Sinkhorn over [n, 4, 4]) inside one vector_section. Up to 18×
  faster than eager PyTorch on small batches (n ≤ 1024).
• sparse_attn is pure vector_section FlashAttention. Per-head
  softmax stats are stored as full [H, D] tiles replicated across D
  to dodge a col-major⇄row-major reshape alias the auto-sync analysis
  can miss. KV is gathered one position at a time via pto.load_scalar
  → dynamic pto.slice_view → pto.load. Beats a hand-written
  torch.gather + npu_fused_infer_attention_score baseline 1.2–1.6×
  across the small/medium shapes typical of this op.

How to build, test and benchmark

See examples/aot/deepseek_v4/OVERVIEW.md.
TL;DR:

# build + correctness-check every kernel
python examples/validate_all_examples.py

# benchmarks
python examples/aot/deepseek_v4/sparse_attn/bench_sparse_attn.py
python examples/aot/deepseek_v4/hc_split_sinkhorn/bench_hc_split_sinkhorn.py

Sample bench results

sparse_attn, vs torch.gather + npu_fused_infer_attention_score (MQA):

  B   M     N    K     pto us     ref us   fused us   pto/ref  pto/fused
------------------------------------------------------------------------
  1   1   128   64     161.15     533.05     265.03     3.31x      1.64x
  1   4   256  128     209.56    1692.93     252.36     8.08x      1.20x
  4   4  1024  128     207.77    6071.60     246.57    29.22x      1.19x
  8   8  2048  128     304.49   24658.49     244.67    80.98x      0.80x

hc_split_sinkhorn, vs eager PyTorch reference:

      n     pto us     ref us  speedup
----------------------------------------
     64     173.27    2803.42   16.18x
   1024     218.70    2761.33   12.63x
  16384    1786.32    2741.09    1.53x