riscv_fpu: don't size-optimize the FP op dispatch

[SolAstrius]

Summary

riscv_emulate_f_opc_op — the interpreter's OP-FP dispatch (fadd/fsub/fmul/fdiv/fsqrt + conversions) — is tagged func_opt_size, which compiles it for size (Oz/minsize) instead of the project default -O2. But the rvjit backend does not emit floating-point yet, so every guest FP instruction is handled here — this "slow path" is actually hot for any FP workload. Dropping the size attribute lets the compiler optimize it normally.

Change

One line: slow_path func_opt_size → slow_path. The slow_path (cold/preserve_most) marking is kept.

Benchmarks (aarch64 host, this dispatch only)

Pure FP-op throughput (tight loops, 40M iters, time-CSR ticks, lower = faster):

op mix	size-opt (before)	normal (after)	speedup
fadd/fsub/fmul	24.8M	11.6M	~2.1×
fdiv	16.1M	6.3M	~2.6×
fsqrt	37.2M	14.0M	~2.7×

Realistic mixed FP (LINPACK 100×100 LU, double):

	before	after
MFLOPS	106	109 (~+3%)

The realistic gain is modest (~3%): in mixed code the integer index/loop/memory ops are JIT-compiled and dominate, so only the FP-handler fraction shrinks. The 2× shows up only when FP-handler time dominates. O3 gives nothing beyond O2.

Size

Function grows ~1.4 KB (riscv_emulate_f_opc_op: 4648 → 6004 bytes); total __text +1.6 KB (+0.67%). On this arm64 build the on-disk binary rounds up by one 16 KB page; on 4 KB-page targets the on-disk delta is correspondingly smaller.

Tradeoff is yours to weigh — modest real-world FP gain for ~1.6 KB of code. Benchmarks (fp-bench, linpack) are bare-metal RVVM firmwares, available on request.