Add logits processor arguments to mlx_lm.generate

[realyxl]

Background

mlx_lm.generate does not currently expose seven logits-processor
parameters that are available in the Python API: logit_bias,
{repetition,presence,frequency}_penalty, and their three _context_size
companions.

These parameters are accepted by
mlx_lm.sample_utils.make_logits_processors(), which produces the list that
generate() / stream_generate() already accept via the
logits_processors keyword argument. mlx_lm.server already exposes the
same seven parameters over HTTP.

Goal

Add the same seven parameters to the mlx_lm.generate CLI.

Motivation

1. Common need in terminal testing. mlx_lm.generate is convenient for
   quickly testing local checkpoints, LoRA adapters, and quantization levels
   from the terminal. These parameters are routinely useful, and sometimes
   necessary, for suppressing the repetition loops that small or
   low-bit-quantized local models often fall into.

2. Consistency.

   • CLI ↔ Python API: generate() already supports
     logits_processors, and mlx_lm.sample_utils.make_logits_processors is
     the standard factory; the mlx_lm.generate CLI does not yet expose them.
   • Sampler ↔ logits-processor: mlx_lm.generate already exposes the
     parallel make_sampler parameters (--temp, --top-p, --top-k,
     --min-p, --xtc-*); the logits-processor side was asymmetric.

3. Minimal and additive. No algorithmic changes. This only wires the
   existing factory into argparse.

4. Why only generate, not chat. mlx_lm.chat might be intentionally
   minimal — it doesn't expose even --top-k. This PR keeps that scope
   unchanged and focuses on mlx_lm.generate, the primary terminal generation
   tool and the one where these knobs are most useful.

Implementation

A single file changed (mlx_lm/generate.py).

• Import make_logits_processors.
• Seven DEFAULT_* constants. None for logit_bias, 0.0 for the three
  penalties, 20 for the three context sizes. This matches the 0.0 == disabled convention used by the sampler defaults in this file and by
  mlx_lm/server.py.
• A small parse_logit_bias argparse type that accepts a JSON object and
  converts keys to int / values to float.
• Seven add_argument calls in setup_arg_parser, ordered to mirror the
  make_logits_processors signature.
• In main(), build logits_processors via make_logits_processors(...)
  (kwarg form) and pass it to generate(...) as the existing
  logits_processors= kwarg.

Tests

No new tests added. pre-commit and the full existing suite
(tests.test_sample_utils + tests.test_generate, 31 tests) pass locally.