[Fix] Gate decode shortcuts and validate dims across layers

[zhiyuan1i]

Summary

Batch of small, independent safety fixes found by auditing the codebase for the same bug classes as recent community fix PRs:

• Decode shortcuts misfire on packed varlen (short_conv.py, ops/rwkv7/fused_recurrent.py, ops/gsa/fused_recurrent.py): shape-equality decode checks (B*T==N, T==1) are wrong when a packed batch has zero-length or multi-token sequences — short_conv misaligns tokens to cache rows, rwkv7 reads/writes out of bounds, GSA treats a packed batch as dense. Gate shortcuts on every sequence having exactly one token.
• Expanded dims not validated at layer construction (7 layers + mom/raven/rwkv7): int(hidden_size * expand) truncates fp-representable integers (50*0.58→28) and silently uses non-integer dims. Round first and reject true non-integers with math.isclose, and assert divisibility for head dims — the convention already used in abc/raven/kda layers. Also fixes raven's false reject of valid expand values (isclose was checked after int-truncation).
• Mamba2 num_heads % n_groups: fail fast instead of a cryptic late reshape error in decode.
• HGRN2 state_size: implement as input_dim * expand_ratio (GLA recurrent state shape).
• raven seqlen max: compute only when RoPE is used (default path skips two device syncs).

Test plan

• Unit tests added/modified: tests/layers/test_layer_dim_validation.py (17 constructor-validation cases, CPU), tests/modules/test_conv.py::test_conv_varlen_decode_detection_with_zero_len_seq
• Dependent tests run: tests/layers/test_layer_dim_validation.py 17/17 pass locally; ruff and compileall clean
• Varlen / CP / model tests: the short_conv test covers a packed varlen batch with a zero-length sequence and cache; GPU-dependent suites run in CI

Benchmark / NCU (kernel changes only)

• N/A — no kernel code changed (validation guards, decode-shortcut gating, one sync-avoidance refactor)

Breaking changes

• Invalid layer configs that were silently truncated or floored now fail loudly at construction; valid configs are bit-identical. The GSA inference shortcut no longer applies to packed varlen inputs (previously wrong for them); dense decode is unchanged.

Checklist

• I have read CONTRIBUTING.md and follow its conventions (code style, docstrings, commit prefixes).
• I have read AGENTS.md and, where my change matches its scope, the relevant skill under .agents/skills.
• Dependent tests pass locally or in CI, and new behavior is covered by tests where applicable (tick as N/A for changes with no testable code, e.g. docs-only).
• Kernel changes include same-hardware before/after benchmark numbers, dense + varlen where applicable (tick as N/A when no kernel code changed).
• This PR is minor/cosmetic-only (typo, formatting, style-only tweaks) — tick only if it is, and justify below.