perf: create the AbortController, drain, and binding Map only when used

[miinhho]

Summary

Closes #25.

Three eager resources every Job paid for without using them are now created on demand, and the state behind them lives in one place.

Cancellation (replaces CancellationBindings) holds one Job's incoming cancellation. After #24 the pieces of "incoming cancellation" had spread across Job: the controller, two pending-source fields, the bindings bag, and ~10 methods (aborted, abortReason, throwIfAborted, isCancellation, isExternalCancellationSource, observeCancellation, recheckCancellation, deadline scheduling inside prepare, controller creation inside cancel). They are now one component with one contract:

• signal — the Job's signal; reading it subscribes to followed sources. The AbortController is created here or by abort(), never before.
• follow(inherited, external) / recheck() / observe() — record, honor without subscribing, subscribe.
• deadline(at) — owns the timer; elapsing cancels the target with a TimeoutError.
• aborted / reason / current / throwIfAborted() / isCancellation(error) — what Job bookkeeping asks.
• dispose() — releases subscriptions and the timer; the signal keeps its state.

Every delivery goes through CancellationTarget.cancel() — the owning Job — so cascades stay the Job's. The component owns no lifetime and no completion state; AGENTS.md's support-state rule is reworded to say exactly that. Job keeps ownership, phases, failures, children, and the result, and holds the component lazily: a plain Job without a seed signal, deadline, or cancellation never creates it.

Two consequences of the consolidation:

• A child's inherited context signal is its parent's by construction, so prepare() no longer reads inherited.signal for it (which used to materialize the parent's controller). Only an explicitly supplied options.context is a foreign source, and one that hands back the owner's signal is still excluded.
• Starting a child observes the parent through parent.cancellation?.observe() rather than by reading parent.signal, so the parent gains a subscription without gaining a controller.

drain() runs only with children. Admission closes when phase becomes "closing", so the check cannot race a late child.

A single-binding ContextFrame holds key/value, not a Map. Frame bindings are an internal Bindings { has, get, keys }; Map satisfies it for several entries, SingleBinding for one. Lookup order and shadowing are unchanged.

No public behavior changes.

Measured

Per-Job microbenchmark (400k iterations, /tmp/job-bench.mjs), packed main (post-#24) vs this branch:

scenario	Node main	Node branch	Bun main	Bun branch
plain Job	0.97 µs	0.82 µs	1.12 µs	0.70 µs
Job + external signal, unobserved	0.97 µs	0.83 µs	1.13 µs	0.77 µs
Job + external signal, signal() read	2.07 µs	2.10 µs	2.00 µs	1.93 µs
HandoffJob + signal, offer/resume	1.35 µs	1.32 µs	1.55 µs	1.24 µs
HandoffJob + signal + values: [provide(...)] (server shape)	1.41 µs	1.36 µs	1.67 µs	1.29 µs

A flat variant that kept the fields on Job measured 1.20 µs on both runtimes for the server shape; the component costs one small allocation per Job that has a seed signal (~0.1 µs). The consolidation is worth that.

Verification

• pnpm format, pnpm check, pnpm build
• pnpm test — 11 files, 131 tests, unchanged
• pnpm pack; server typecheck, vitest --project @tiberjs/server (12/12), leak-probe (live 0, graph 0) against the packed tarball