Add GPTQ quantized linear layer support for Qwen2 (#3650)

[astorise]

Summary

Adds support for loading GPTQ-quantized checkpoints, as requested in #3650.

• candle-gptq-kernels: a new crate with fused dequantize+GEMM CUDA and
  Metal kernels for GPTQ 4-bit linear layers (CUDA kernel includes a
  tensor-core path adapted from the Marlin project — see
  candle-gptq-kernels/kernels/marlin/ATTRIBUTION.md for license/provenance).
  Kept as its own crate rather than folded into candle-kernels so it's only
  built for checkpoints that actually use GPTQ; see the crate's README for the
  full rationale.
• candle_transformers::quantized_gptq: portable CPU dequantize-at-load path
  (gptq_linear/GptqConfig) that works on any backend without requiring the
  gptq-cuda/gptq-metal feature.
• candle_transformers::models::gptq_qwen2: a Qwen2 model that routes every
  attention/MLP projection through the GPTQ path above.
• examples/quantized-gptq-qwen2: end-to-end example that downloads a GPTQ
  checkpoint from the Hub (default
  Qwen/Qwen2-0.5B-Instruct-GPTQ-Int4)
  and runs text generation through it.

Why split into multiple PRs

This is the first of three PRs addressing #3650 (GPTQ, then AWQ, then
block-wise FP8). The original branch implemented all three formats together,
but several commits ended up mixing CUDA/Metal/CI work across formats, which
made the diff hard to review and hard for CI to validate per-format. This PR
is the GPTQ-only slice, reconstructed to be independently buildable and
testable so reviewers/CI can validate GPTQ in isolation before the AWQ and
FP8 PRs land on top of it. The two follow-up PRs:

• AWQ support + a QuantizedLinear/QuantMethod abstraction unifying
  GPTQ/AWQ behind one Module impl, generalizing this PR's example into a
  single quantized-qwen2 example that auto-detects the format from
  config.json.
• Block-wise FP8 (DeepSeek-V3-style) support, extending the same
  QuantizedLinear enum with an Fp8 variant.

Test plan

• cargo test -p candle-transformers --lib quantized_gptq — dequantize
  roundtrip unit test
• cargo test --manifest-path candle-gptq-kernels/Cargo.toml --no-default-features — kernel crate builds/tests standalone (CPU)
• cargo clippy -p candle-transformers -p candle-examples --example quantized-gptq-qwen2 --all-targets -- -D warnings
• cargo fmt --check
• Manually ran cargo run --example quantized-gptq-qwen2 --release against
  Qwen/Qwen2-0.5B-Instruct-GPTQ-Int4
• CI: cargo test -p candle-transformers --features cuda,gptq-cuda and
  cargo test --manifest-path candle-gptq-kernels/Cargo.toml (tensor-core
  kernel correctness) on CUDA; --features metal GPTQ kernel test on
  Metal — added as new steps in ci_cuda.yaml/ci_metal.yaml.

Closes #3650 (GPTQ portion; AWQ and FP8 to follow in separate PRs).