Fix staging cold-start timeout

[m-messer]

Staging returned 503/500 on every request: the worker's cold-start import (torch + now pydantic/pydantic-core from the lf_toolkit v1.1.1 bump) takes ~34s on the 1024MB Lambda, longer than shimmy's 30s worker-send-timeout, so shimmy killed the half-booted worker before it became ready and never recovered.

• evaluation.py / models/init.py: dispatch models via importlib against an AVAILABLE_MODELS allowlist instead of eagerly importing every model module. torch is no longer pulled in at worker startup, only when a model that needs it actually runs. Cold import drops ~34s -> ~8s (measured CPU/mem-constrained to match the Lambda).
• models/utils.py: build the torch NeuralLM class lazily so importing utils for csv_to_lists / shard_for (shannon_letters_ngram, shannon_words_ngram) does not pull in torch either.
• Dockerfile: raise FUNCTION_WORKER_START_TIMEOUT / FUNCTION_WORKER_SEND_TIMEOUT to 150s (< the 175s Lambda timeout) so shimmy waits out a slow first request instead of discarding the worker; make FUNCTION_RPC_TRANSPORT=stdio explicit.

Verified in AWS RIE Lambda mode, container constrained to 1024MB / 0.6 vCPU: cold torch-free model 13.9s -> 200; cold basic_nn (first torch load) 35s -> 200 on direct invoke; warm calls ~0.2s. Worker now survives the slow first call. pytest + flake8 pass.

Claude-Session: https://claude.ai/code/session_01UsgxJUsq1vG176rbznHRqz