Use lazy/no-FPU context switching when safe

[drpaneas]

Summary

The runtime always uses the full __go_swapcontext path, which saves and restores FPU state on every goroutine context switch even when neither goroutine uses floating-point operations.

Why

• Integer-only goroutines currently pay avoidable context-switch cost.
• runtime_sh4_minimal.S already contains __go_swapcontext_lazy and __go_swapcontext_nofpu, but the scheduler never selects them.
• On a single-core Dreamcast runtime, context-switch overhead directly affects frame budget and goroutine-heavy workloads.

Evidence

• runtime/runtime_sh4_minimal.S documents the current full path at about 88 cycles with FPU save/restore included.
• The same file documents __go_swapcontext_lazy as saving about 19-25 cycles per skipped direction and __go_swapcontext_nofpu as saving about 50 cycles versus the full path.
• The scheduler currently uses only __go_swapcontext and does not track per-goroutine FPU usage or pass any FPU flags.

Direction

• Investigate a sound way to track whether a goroutine has used floating point.
• Teach the scheduler to select among __go_swapcontext, __go_swapcontext_lazy, and __go_swapcontext_nofpu when it is safe to do so.
• Prefer correctness first: any optimization must preserve FPU state across goroutines that do use floating point.
• Benchmark the real savings on hardware for integer-heavy workloads.