Add Gear model support

[crinex]

Summary

This PR adds initial mlx-lm support for the Gear causal language model
architecture.

Included scope:

• new mlx_lm.models.gear architecture module for model_type="gear"
• full and sliding attention layers with Gear Q/K normalization
• local conv_mixer layers with depthwise key/value convolution state
• Gear RMSNorm semantics using (1 + weight)
• scaled token embeddings matching the reference implementation
• HF Conv1d weight layout sanitization for key_conv and value_conv
• model smoke coverage in tests/test_models.py

Excluded from PR scope:

• local validation scripts and model artifacts
• bidirectional Gear attention mode
• model-specific training examples

Validation

Local model artifact:

• sft_final
• HF architecture: GearForCausalLM
• model_type: gear

Weight key check against the local safetensors header:

actual_count=362 expected_count=362
missing=0 extra=0

Commands run:

.venv-contrib/bin/python -m py_compile mlx_lm/models/gear.py tests/test_models.py
git diff --check
.venv-contrib/bin/pre-commit run --files mlx_lm/models/gear.py tests/test_models.py
.venv-contrib/bin/python -m unittest tests.test_models.TestModels.test_gear tests.test_evaluate tests.test_tokenizers.TestTokenizers.test_tokenizers
.venv-contrib/bin/python -m unittest discover tests/
.venv-contrib/bin/python - <<'PY'
from pathlib import Path
from mlx_lm.utils import load_model

model, config = load_model(Path('../contribution-prep/models/sft_final'), lazy=True)
print('model_type', model.model_type)
print('layers', len(model.layers))
print('vocab_size', config['vocab_size'])
print('loaded_lazy', True)
PY
.venv-contrib/bin/mlx_lm.generate \
  --model ../contribution-prep/models/sft_final \
  --prompt "What is 2+2?" \
  --max-tokens 16 \
  --trust-remote-code
.venv-contrib/bin/mlx_lm.convert \
  --hf-path ../contribution-prep/models/sft_final \
  --mlx-path /private/tmp/gear_mlx_q4_smoke_20260714_1635 \
  --quantize \
  --q-bits 4 \
  --q-group-size 64 \
  --trust-remote-code
.venv-contrib/bin/mlx_lm.generate \
  --model /private/tmp/gear_mlx_q4_smoke_20260714_1635 \
  --prompt "What is 2+2?" \
  --max-tokens 16 \
  --trust-remote-code

Result:

black....................................................................Passed
isort....................................................................Passed
Ran 3 tests in 8.347s
OK
Ran 210 tests in 41.611s
OK (skipped=1)
model_type gear
layers 26
vocab_size 125184
loaded_lazy True
Prompt: 18 tokens
Generation: 16 tokens
Peak memory: 1.720 GB
Quantized model with 4.502 bits per weight.
Quantized generation peak memory: 0.508 GB

Environment note:

• Validation used a dedicated .venv-contrib environment with mlx==0.32.0,
  transformers==5.13.1, lm-eval==0.4.12, and datasets==5.0.0.
• MLX runtime checks were run outside the sandbox because the sandbox cannot
  access the local Metal device.
• The local sft_final artifact requires --trust-remote-code for
  tokenizer/config loading because it contains auto_map custom code.