Count DSv4 attention FLOPs in HybridModel training metrics

[FDecaYed]

• I, the PR author, have personally reviewed every line of this PR.

What does this PR do?

The current HybridModel FLOPs calculation omits C/H/W attention layers. Count their query/KV/output projections, window attention, compression, ratio-4 indexer and sparse attention, and ratio-128 attention. Read layer counts and MTP depth from the actual hybrid pattern, and use DSv4's head dimension rather than an uninferred MLA argument default.

Reuse main's real-token and squared-sequence-length statistics for packed batches. Packed Top-K saturation remains an estimate based on token-weighted mean sequence length; the existing statistics do not retain individual document lengths. Keep the existing forward/backward FLOPs convention and include MTP normalization/projection work.

Related to #5795. Adapted from the DSv4 accounting in #5142, #5358, and #5762. Independent of #7336 and #7469. This changes reported FLOPs, not model execution speed.

Validation

• 48 CPU cases passed: the 40 existing calculator cases plus 8 focused DSv4 cases covering W/C/H, packed statistics, repeated MTP, and dense CSA without an indexer.
• The CPU harness executes the production calculator and pattern parser with only unrelated package imports isolated. The new W case fails against the original calculator as expected.
• Required autoformat checks, targeted Ruff, syntax compilation, and diff whitespace checks passed. The separately inspected training.py Pylint output has no diagnostics on changed lines; that file has existing diagnostics outside this change. Advisory mypy lacks local dependency stubs.
• No GPU is required to validate these arithmetic changes; full training CI has not run for this draft.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[FDecaYed]

/ok to test 4734000