…ntext

Addresses five review comments on the TurboModule crash-time context PR
(#6227):

- Pin the Scope on each tracker frame at push time. `popTurboModuleCall`
  and `relabelTurboModuleCallKind` now always operate against the scope
  captured at push, so an async call that spans a scope switch
  (`withScope`, isolation-scope swaps) clears the right scope instead of
  leaking stale `turbo_module` context onto the original. `pop` no
  longer takes a `scope` argument — it was a footgun.

- Replace `setTag(key, undefined)` on clear with an empty-string
  sentinel. The native `setTag(key: string, value: string)` TurboModule
  spec requires a string; `undefined` would either be rejected or
  silently dropped at the bridge. `setContext(key, null)` remains the
  canonical "no active call" signal; the empty tags exist only to scrub
  stale `name`/`method` from the previous call.

- Defer `wrappedModules.add(module)` until after at least one method
  has been successfully wrapped. Previously, a JSI HostObject whose
  methods materialise lazily was permanently locked out after the
  first (empty) discovery pass — subsequent calls short-circuited even
  though the module had since gained methods.

- Strengthen the sealed-module test to spy on `pushTurboModuleCall`
  and assert exactly one push per real call. The previous assertion
  only checked the post-call stack was empty, which is also true under
  double-wrapping (each wrapper pushes and pops symmetrically).

- Move the changelog entry from the already-released `## 8.13.0`
  section into `## Unreleased`.