runtime: block signal_recv on targets without signal delivery (#5619)

[neomantra]

This was worked through with LLM. The text below is LLM-generated and I have read and reviewed all the code.

But I did finally get the BubbleTea List demo running in browser, compiled by TinyGo!

---

Summary

Make the stubbed os/signal.signal_recv block forever instead of returning
immediately, so signal.Notify no longer starves cooperative schedulers on
wasm and baremetal targets.

Fixes #5619

Problem

src/runtime/signalstub.go (build tags tinygo.wasm || baremetal) stubbed
signal_recv as return ^uint32(0). Upstream's os/signal.loop calls
signal_recv in a tight loop with no yield point — it relies on the
runtime blocking until a signal arrives, as the POSIX implementation in
runtime_unix.go does. With the immediate-return stub, the watcher
goroutine started by signal.Notify spins forever.

On a cooperative scheduler a spinning goroutine is always runnable, so the
scheduler never idles: on wasm, _start never returns to the host event
loop and the browser tab (or node/wazero) pins a core with all other
goroutines starved. Any program calling signal.Notify is affected —
Bubble Tea does so unconditionally, which is how this surfaced (a V8
profile of the hung program showed 99.7% of ticks in os/signal.loop).

Fix

signal_recv now calls deadlock() — the same primitive a blocking empty
select uses — parking the watcher goroutine forever. That matches the
real implementation's observable behavior on a system where no signal ever
arrives: Notify succeeds, the channel simply never receives anything, and
everything else keeps running. This is also what gc's js/wasm port does.

Verification

• New behavioral test testdata/signalnotify.go (signal.Notify, then
  time.Sleep, then print done), registered for all platforms. The
  existing signal.go test is skipped on wasm/baremetal/windows, which is
  why this had no coverage.
• Red: on stock dev @ 86d58db the wasm test hangs until the Go test
  timeout kills it.
• Green with this change: TestBuild/WebAssembly/signalnotify.go and
  TestBuild/Host/signalnotify.go both pass (the host run exercises the
  real POSIX signal_recv path, guarding both implementations).
• End to end: a go-booba-adapted Bubble Tea bubbles/list app compiled for
  GOOS=js GOARCH=wasm previously froze the browser tab at 100% CPU right
  after startup; with this change it runs interactively.
• deadlock() is defined by all four scheduler implementations
  (cooperative, threads, cores, none), covering the stub's whole build-tag
  surface.
• Not run locally: cortex-m-qemu / AVR / RISC-V builds of the new test
  (this environment lacks the LLVM source checkout for compiler-rt); CI
  covers those. If os/signal turns out not to fit AVR flash, the test can
  be excluded for AVR the way json.go/stdlib.go are.

Context

Found while verifying the large-parameter-spilling work (#5615) in a real
browser; it is independent of that change and reproduces on stock dev.