aarch64: gate FP16 vector helpers with target_feature

[kpernyer]

What changed

Adds #[target_feature(enable = "fp16")] to the four aarch64 NEON vector FP16 inline-assembly helpers in gemm-common:

• vmulq_f16
• vaddq_f16
• vfmaq_f16
• vfmaq_laneq_f16

Why

These helpers emit Full FP16 instructions but were not target-feature-gated, unlike the scalar FP16 helpers nearby. Generic aarch64-unknown-linux-gnu code generation therefore rejects them with instruction requires: fullfp16.

Keeping the feature requirement on the functions preserves GEMM's runtime FP16 dispatch. Globally enabling +fp16 is not portable because runtime feature detection becomes compile-time true for that build.

Fixes #46.

Validation

• cargo fmt --check
• CARGO_HOME=/tmp/gemm-fp16-cargo-home CARGO_TARGET_DIR=/tmp/gemm-fp16-target-generic cargo test --package gemm-common --package gemm-f16

The targeted test command was run on Linux aarch64 with a clean Cargo configuration, so the machine-wide +fp16 override was not in effect. cargo test --workspace was also attempted but is currently blocked before tests by a missing system fontconfig development package (fontconfig.pc) required by an unrelated dependency.