Packed THD training (offline-packed LLM SFT and VLM in-batch packing)
over-counts attention FLOPS by treating the whole pack as one length-
seq_length sequence (pack_length²). Actual attention work is Σᵢ sᵢ²
over the real sub-sequence lengths.

New helper accumulate_flops_metadata() in flop_utils.py extracts the
real sub-seq lengths from cu_seqlens (preferring cu_seqlens_unpadded
when present) and feeds Σᵢ sᵢ² into the existing seqlen_squared_sum
accumulator from NVIDIA-NeMo#3529. Falls back to BSHD mbs * seq_len² when no
cu_seqlens is provided — bit-exact identical to legacy on dense
pretraining and non-packed paths.

Wired into gpt_step, vlm_step, qwen3_vl_step, and qwen3_omni_step.

Verified on cw-dfw (same seed, same data, same iter times, identical
loss values across paired runs — only the reported TFLOPS differs):
- qwen3_8b_sft  seq=2048: baseline 162.6 vs fix 155.8 TFLOP/s/GPU (+4%)
- qwen3_8b_sft  seq=4096: baseline 339.9 vs fix 156.7 TFLOP/s/GPU (+117%)
- qwen35_vl_9b_sft     : baseline 261.6 vs fix  88.9 TFLOP/s/GPU (+194%)

The seq=2048→4096 pair on the same LLM recipe is the cleanest
demonstration: the fix is near-flat (155.8 vs 156.7 — attention work
is determined by per-sample lengths, not pack length) while the
baseline doubles because its pack_length² scales quadratically.

9 new unit tests in test_flop_utils.py::TestAccumulateFlopsMetadata
cover the BSHD fallback, THD with cu_seqlens, padded cu_seqlens via
cu_seqlens_argmin, cu_seqlens_unpadded precedence, additive
accumulation, and the regression headline (32-sample pack → 32x
smaller attention work than BSHD approximation).

Signed-off-by: Chen Cui <chcui@nvidia.com>