fix(sampling): apply presence and frequency penalties

[taking-lying-flat]

Chat and text completion requests accept presence_penalty and frequency_penalty, but currently drop both values before sampling. This change carries them through SamplingParams and applies logit -= frequency * count + presence * (count > 0) before greedy selection or temperature/top-k/top-p sampling.

Counts belong to each request and include generated tokens only. They are updated on the sampling stream, so overlapping execution sees the latest sampled token before its CPU acknowledgement arrives. Intermediate prefill chunks do not contribute counts; requests with zero penalties allocate no counter. The HTTP models reject non-finite penalty values to prevent NaN logits.

Validation: 181 checks passed: 73 existing API, wire-format, accounting and scheduler checks, plus 108 local HTTP/CUDA checks for parameter forwarding, positive/negative penalties, final-token handling, mixed greedy batches, changing batch membership, stream ordering, non-finite inputs and zero-penalty parity with the previous sampler. The post-submission checks also cover vocabulary sizes 128,256 and 248,320, batch sizes 1/8/32, and FP16/BF16/FP32 logits. This branch contains production code only; the additional checks were run from files outside the repository.

Environment: Intel Core i7-12800HX, NVIDIA RTX A1000 Laptop GPU (4 GiB), driver 595.84, PyTorch 2.13.0+cu130, Triton 3.7.1. GPU checks used synthetic logits; no model checkpoint or full-model/multi-GPU inference was run. The installed Torch/Triton versions differ from the repository's pinned ranges.

Exact validation commands, run from the repository root (163 checks, then 18 additional checks):

/tmp/freetoken-497-venv/bin/python -m pytest -q -o pythonpath=python \
  /tmp/freetoken-penalty-artifacts/test_penalty_review.py \
  /tmp/freetoken-penalty-artifacts/test_sampling_penalties.py \
  tests/server/test_openai_api.py \
  tests/server/test_effort_dialect.py \
  tests/server/test_message_wire.py \
  tests/server/test_generation_accounting.py \
  tests/scheduler/test_req_append_host.py \
  tests/scheduler/test_scheduler_chunked_prefill.py \
  --maxfail=2 --tb=short -o faulthandler_timeout=30

/tmp/freetoken-497-venv/bin/python -m pytest -q \
  /tmp/freetoken-penalty-artifacts/test_post_pr_review.py \
  --maxfail=1 --tb=short -o faulthandler_timeout=30