[ruocheng-nv]

Description

C++/CMake part of #20878 (GLM-5.3-Flash FP8 on Hopper). The Python changes are in #21179 and the
Triton FP8-activation MoE is in #21250. These changes are model-agnostic and do not depend on either.

Build FlashMLA without relocatable device code

• flash_mla_cpp_tllm was built with CUDA_SEPARABLE_COMPILATION ON (nvcc -rdc=true). Nothing in
  FlashMLA needs cross-TU device linking (its own setup.py builds without -rdc), and -rdc makes
  ptxas spill in the attention kernels:
  • sm90 sparse prefill (d_qk 512): 0 -> 2044 B spill stores, 296 B stack
  • sm90 sparse fp8 decode (h64): 0 -> 220 B
  • sm100 sparse prefill (h128): 0 -> 44 B
  • sm100 sparse decode (h64): 0 -> 12 B
• H200, flash_mla_sparse_fwd with 13312 queries x 64 heads x 2176 top-k rows (d_qk 512):
  17.07 -> 7.85 ms per call, outputs bit-identical. Same FlashMLA commit and nvcc; only -rdc differs.

FP8 block-scale CUTLASS MoE: row expansion and activation

• expandInputRowsFp8BlockScaleKernel binary-searched the expert offsets for every permuted row.
  Each block now takes a contiguous run of permuted rows, searches once for the first row and
  advances the expert linearly (rows are sorted by expert).
• doActivation always launched 256-thread blocks with eight channels per thread, so a 512-wide
  intermediate slice (tensor parallelism) left three quarters of each block idle. The block is now
  sized to the row width in whole warps.
• GLM-5.3-Flash FP8, H200 TP4, 13K-token prefill (45 MoE layers, 106K expanded rows): row
  expansion 29.6 -> 18.9 ms, activation 13.7 -> 6.6 ms; C1 TTFT 522 -> 507 ms. GSM8K unchanged.

Measurements were taken on the #20878 branch (older main); this branch is the same two commits
cherry-picked onto current main without conflicts.

Test Coverage

• Existing FlashMLA sparse attention and CUTLASS FP8 block-scale MoE tests (outputs bit-identical
  for the FlashMLA change; the MoE kernels change scheduling only).
• End-to-end GSM8K on GLM-5.3-Flash FP8 (H200 TP4) unchanged.

PR Checklist

• PR description clearly explains what and why.
• No API changes.