perf(coro): allocate panic snapshots lazily

[cpunion]

Move the bounded 64-PC panic snapshot out of every runtime G and allocate it only when a managed snapshot is requested. Stackless language panics retain their scheduler-owned logical trace and pay no allocation.

The legacy SA_SIGINFO path remains allocation-free: it reuses an existing G-owned store or a static emergency store matching the already process-global fault-source semantics. Lifecycle paths release owned stores, and a source gate prevents the signal callback from reaching the allocating entry.

Measured at 10,000 parked Gs with GOMAXPROCS=1:

• parent c773ea5: 40,779,776 bytes RSS
• candidate: 32,751,616 bytes RSS (-19.7%)
• Go gc: 31,539,200 bytes RSS
• candidate incremental slope: 2,502 bytes/G vs Go 2,798 bytes/G

Validation:

• cd runtime && go test ./...
• go test ./cl -run TestCoro(ExplicitStatusRecoverIRNativeAndWasm32|ExplicitStatusPanicNativeAndWasm32)$ -count=1
• go test ./internal/build -run TestCoroExplicitPanicNativeNoStdlibRuntimeE2E -count=1
• native core workload build/run at 0, 1k, 5k, and 10k parked Gs
• WASI core workload build and Wasmtime execution

The legacy PCLN signal integration command is blocked before its selected subtest by the same existing pclntab_external.go uintptr-retention audit on both this branch and exact parent c773ea5.