fix: annotate aarch64 fp16 vector helpers with target features

[K12231217]

Summary

Annotate the four aarch64 helpers that emit 8-lane FP16 vector instructions with the function-level fp16 target feature:

• vmulq_f16
• vaddq_f16
• vfmaq_f16
• vfmaq_laneq_f16

The existing microkernel dispatch remains unchanged. This does not enable FP16 globally or raise the baseline CPU requirement for callers that do not enter the FP16-dispatched path.

Reproduction

The standalone workspace uses dev opt-level 3, which hides the issue. A top-level downstream workspace using an unoptimized debug profile reproduces it without additional CPU flags:

env -u RUSTFLAGS CARGO_PROFILE_DEV_OPT_LEVEL=0 CARGO_PROFILE_DEV_DEBUG=1 cargo build -p gemm-f16 --target aarch64-unknown-linux-gnu

Before this change, LLVM rejects the instantiated vfmaq_laneq_f16 assembly with instruction requires: fullfp16.

Verification

• The reproduction command passes after the change using a fresh target directory.
• env -u RUSTFLAGS cargo build -p gemm-f16 --release --target aarch64-unknown-linux-gnu
• cargo fmt --all -- --check
• cargo test --workspace: 5 tests passed; all doctests passed.

Tested from the gemm-common 0.18.2 release baseline because that is the downstream dependency version exposing the failure.