Respect TensorRT-LLM input limits during generation

[sdivyanshu90]

TRT-LLM generation currently borrows unused output capacity for prompts, so an engine configured with max_input_len=2048 can receive 2304 input tokens. Include the separate engine input cap in the truncation budget while retaining the total sequence cap, requested generation length, and truncation direction.

Add CPU tests using the real constructor and generate_until with a recording engine, covering all three truncation directions, short/boundary/long prompts, and per-request generation lengths. Include the mechanical typing, formatting, and exception-preserving corrections required by pinned lint checks on the touched backend.

Fixes #4406.

Validation:

• Before the fix: 5 new regression cases failed; 10 controls passed.
• python -m pytest -q tests/models/test_trtllm_input_limit.py tests/models/test_model_utils.py: 54 passed.
• pre-commit run --files lm_eval/models/trtllm_causallms.py tests/models/test_trtllm_input_limit.py: all applicable pinned hooks passed.
• The exact issue reproduction changes from 2304 to 2048 input tokens.
• Actual TensorRT/GPU inference was not exercised.

[sdivyanshu90]

Hi @baberabb, could you please review this PR whenever you have some time. Thanks!