perf(vm): make PrepareCall allocation-free for native fold loops

[nnunley]

Fixes the allocation regression from #727 and restores a green make bench-ratchet on main. Bisect and numbers in #791.

#727 routed reduce and some through PrepareCall, which heap-allocated a PreparedCall and an argument slice per invocation and resolved the target through resolveBytecodeCall, which packs a rest list for variadic fns before the preparation could reject them. The IR compiler performs thousands of short reductions per compile, so BenchmarkIRCompile gained ~1,900 allocs/op, over the ratchet's 2% deterministic bar.

• PrepareCallInto(p, fn, arity) fills a caller-owned PreparedCall, so the hot loops keep it on their stack.
• Argument slots live in a new Frame.prepArgs that survives pool reuse and does not alias argbuf, which installBytecodeCall overwrites on a same-frame tail call.
• unwrapBytecodeFn, split out of resolveBytecodeCall, inspects the target without allocating, so variadic and wrong-arity fns are rejected before any packing. PrepareCall remains as the allocating form.
• TestPrepareCallIntoDoesNotAllocate pins the zero-allocation contract; the existing PreparedCall tests (tail-call rebind, error reuse, arity mismatch) still pass.

On an Apple M3, plugged in, BenchmarkIRCompile [bytecode]:

	allocs/op	bytes/op	ns/op
#780 baseline	114,255	4,963,229	15.05M
main (bdd8268)	116,149	5,080,290	17.01M
this PR	113,722	4,935,179	15.48M

make bench-ratchet on this tree: deterministic OK, anchor +0.7%, IRCompile [bytecode] +2.0% ok, InitFromLGB -2.4%, IRCompile [gogen_ir] -10.2%, 0 regressions. go test ./... -skip TestClojureTestSuite and go test ./test/e2e pass; make check-generated clean (manifest and sums refreshed).

[mparrett]

Reviewed head 1b4d8f0. No findings.

I checked the PreparedCall lifecycle and frame-pool ownership paths, including same-frame tail-call rebinding and rejection of unsupported targets. go test ./pkg/vm ./pkg/rt -count=1, the corresponding race run, 100 repetitions of the PreparedCall tests, make check-generated, and git diff --check all pass.

Independent base/head allocation checks reproduced the intended improvement. On the ordinary bytecode path, steady BenchmarkIRCompile samples dropped by 2,427 allocs/op and about 145 KB/op. With generated IR enabled, steady samples also dropped by about 1,486 allocs/op and 55 KB/op. I did not use wall-clock results as review evidence because this host was under heavy unrelated load.

The current merge conflict appears limited to generated.manifest / generated.sums; rebasing and regenerating those files should be the remaining integration step.