…or FLOPs calculator

Closes coverage gaps in tests/unit_tests/training/utils/test_flop_utils.py:
the existing suite covered MoE, hybrid (Mamba/GDN/MTP), SWA, and attention
output gate, but the transformer_flops paths for Multi-Latent Attention
(DeepSeek-V2/V3), explicit mtp_num_layers, and the model-provider override
were untested. Functional tests cover only Llama / Qwen3-MoE today, so
DeepSeek-V3 — the most active model family in this repo — has had no
TFLOPS regression coverage at all.

Adds 10 unit tests across 4 new classes (337 lines, no production code):

- TestMLAFlops (6 tests)
  - test_mla_with_q_lora_exact_formula: bit-exact match against the
    closed-form MLA FLOPs (DeepSeek-V3 Q-LoRA + KV-LoRA path)
  - test_mla_without_q_lora_exact_formula: bit-exact match for the
    no-Q-compression branch (h * n_heads * (qk + qk_pos))
  - test_q_lora_reduces_q_projection_flops: invariant — Q-LoRA
    compression must reduce attention FLOPs when rank is small enough
  - test_mla_differs_from_standard_attention: MLA != MHA path
  - test_mla_batch_size_scales_linearly: f(B=4) == 4 * f(B=1)
  - test_mla_seq_length_quadratic_growth: doubling s grows total > 2x
    (because core attention is O(s^2))

- TestMLAWithMoE (1 test)
  - test_mla_moe_combination_positive_and_distinct: DeepSeek-V3 shape
    (MLA + MoE) produces sensible totals distinct from MHA + MoE

- TestExplicitMtpInTransformerPath (2 tests)
  - test_explicit_mtp_increases_flops: directional invariant
  - test_explicit_mtp_exact_delta: bit-exact delta matching the
    per-layer MLP, per-layer attention, MTP norms/proj, and (mtp+1)
    logit-scaling contributions

- TestProviderOverride (1 test)
  - test_provider_override_short_circuits: when the model exposes
    _get_num_floating_point_operations, the calculator is bypassed
    and the override is invoked exactly with the requested batch_size

The MLA arithmetic mirrors the closed form in flop_utils.py
(num_floating_point_operations -> transformer_flops, lines 343-398),
so any future regression in that branch surfaces immediately.

MockModelConfig is extended with q_lora_rank, kv_lora_rank, qk_head_dim,
qk_pos_emb_head_dim, v_head_dim — all defaulted so existing tests are
unaffected.

Signed-off-by: lonexreb <[email protected]>