feat(glm53): add text runtime and measured performance paths

[ro99]

Summary

• add text-only GLM-5.3-Flash support to chat and the OpenAI-compatible server
• implement checkpoint validation, hybrid KDA/MLA execution, W8A8 FP8 kernels, fused KDA and mHC transitions
• add dynamic host-FP8 expert execution, topology-aware pipeline/TP2 placement, and page-level expert reuse
• add MTP verification, continuous iteration scheduling, paged sequence state, and prefix caching
• reject vision inputs and preserve exact text behavior without hybrid cache quantization

Measured result

On the available dual RTX 3090 CUDA pair (--devices 0,1):

• load: 1.65 s
• prefill: 0.80 tok/s on the 13-token chat prompt
• decode: 0.90 tok/s greedy

The mixed RTX 3090 / RTX 5060 Ti pair measured 0.91 tok/s decode. Nsight attribution reduced routed-expert prefill H2D traffic from 70.34 GB to 14.04 GB.

Validation

• cmake --build build-cuda -j $(nproc)
• build-cuda/strata-tests
• 331 tests passed, 38 skipped
• deterministic text matched between the measured device configurations

Deliberate defaults

• MTP is implemented and available through STRATA_GLM53_MTP=1, but remains opt-in until acceptance amortizes verification for the workload.
• absorbed resident MLA remains experimental and opt-in until its BF16 boundary equivalence gate is closed.
• vision and hybrid cache quantization are out of scope.