speculative : fix DFlash 0% draft acceptance on Metal (sanitize non-finite target features)

[apolinario]

DFlash speculative decoding on Metal gets 0% draft acceptance (all draft logits NaN), making it slower than base decoding. Sanity checking, the same setup works if the inference is done purely on CPU. The issue seems to be that the Metal mat-mat kernels stage f32 activations as f16, and Laguna's attention-sink activations (~±9e5 in the late residual) overflow to Inf in the hidden states DFlash extracts for the drafter. One poisoned row NaNs the drafter's KV cache from round one. Base generation is unaffected since only the last row gets sampled.

This PR fixes that by clamping non-finite values in the extracted features before the DFlash encoder (NaN→0, ±Inf→±f16 max). On an M4 Max with Laguna-S-2.1 Q4_K_M + the BF16 drafter (greedy, draft-n 7): acceptance 0%→54%, decode 35→49.5 tok/s (~1.4x base).

Repro (Apple Silicon, ≥96 GB):

./build/bin/llama-cli -m laguna-s-2.1-Q4_K_M.gguf -md laguna-s-2.1-DFlash-BF16.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7 -c 8192 -st -fa on --temp 0 --top-k 1 \
  -p "def fibonacci(n):" -n 96 --no-warmup -v 2>&1 | grep "draft acceptance"
# draft acceptance = 0.00000 (0 accepted / 637 generated)
# with -ngl 0 -ngld 0 (CPU): acceptance ~0.29
# with -ub 8 (prefill via mat-vec kernels, f32 reads): acceptance recovers

[joerowell]

Thank you for your contribution! Don't know how we missed that.

[usofrob]

When I compile for Vulkan I get this error:
common/speculative.cpp:1091:35: error: ‘isfinite’ is not a member of ‘std’
1091 | if (!std::isfinite(v)) {

Compile command:
cmake -B build -DGGML_VULKAN=ON
cmake --build build --config Release -- -j 24

Is there something different I need to do?
It worked before I took in this change.
I'm compiling on Ubuntu 24.04.

EDIT TO ADD: This issue isn't happening on the master branch now. So, please disregard.