fix(quant): resolve dtype mismatch casting F8E4M3 to BF16 in UnquantLinear (#2072)

[glaziermag]

Fixes Issue #2072.

This PR aims to resolve a dtype mismatch panic that occurs when UnquantLinear attempts to dynamically cast weights from FP8-quantized tensors (DType::F8E4M3) natively. Since candle_core::DType::to_dtype does not export PTX CUDA instructions for F8E4M3 -> BF16 standard resolutions yet, this branch explicitly maps the casting operation via the scalar_fp8::ops::fp8_to_dtype bridge instead.

I've run some local verification tests on an L4 GPU (g2-standard-32) to make sure this doesn't break existing UnquantLinear fallback logic for native models (like Llama-3 in BF16).

1. Blast Radius Assessment

• Component: mistralrs-quant/src/unquantized/mod.rs (UnquantLinear::forward).
• Potential Impact: General unquantized LLM inference operations loading in standalone precision.
• Verification: Tests confirm UnquantLinear fallback algebra does not panic on baseline tests for unquantized models.

2. Execution Logs (Before vs. After)

Before logs (reproducing the issue):
$ cargo run --release --features cuda -- serve --port 1234 -m neuralmagic/Meta-Llama-3-8B-Instruct-FP8

2026-04-10T02:03:31.595045Z  INFO mistralrs_core::models::llama: Using fp8 quantization: 8 bits.
2026-04-10T02:04:13.296123Z ERROR mistralrs_core::engine: step - Model failed with error: unexpected dtype, expected: BF16, got: F8E4M3

After logs (with this patch applied):
$ cargo run --release --features cuda -- serve --port 1234 -m neuralmagic/Meta-Llama-3-8B-Instruct-FP8

2026-04-10T02:15:16.111630Z  INFO mistralrs_server_core::mistralrs_for_server_builder: Model loaded.
2026-04-10T02:15:16.112250Z  INFO mistralrs_core: Beginning dummy run.
2026-04-10T02:15:24.210607Z  INFO mistralrs_core: Dummy run completed in 8.098182927s.
2026-04-10T02:15:24.214426Z  INFO mistralrs::commands::serve: Server listening on http://0.0.0.0:1234

3. Regression Tests Run

Ran the standard mistralrs-quant test suite natively with CUDA to help verify that the native operations remain structurally intact over BF16.
$ cargo test --release --features cuda --package mistralrs-quant

...
test fp8::quantize::tests::test_cublaslt_matmul ... ok
test hqq::quantize::test::test_quantize_hqq ... ok
test cublaslt::tests::test_fused_batch_matmul_f8e4m3_nobias ... ok
test cublaslt::tests::test_fused_batch_matmul_f8e4m3_out_bf16 ... ok
test blockwise_fp8::ops::tests::test_blockwise_fp8_gemm ... ok

test result: ok. 41 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 8.96s

I'd appreciate any feedback on this approach! Thank you for reviewing!