perf(stages)!: build the level chain once at freeze

[illia1f]

A level used to be an object built per next call: N stages meant N allocations, and a retry re-allocated everything below it. A level holds nothing belonging to one call, so it never needed building per call.

Each plan now builds a chain of LevelEntry<TResponse> delegates once, when the dispatch map freezes. The request, the provider, and the token reach a level as arguments, so one chain serves every dispatch of its request type.

• Per-dispatch allocation no longer grows with the chain. Five stages cost what no stages cost, and a repeated next call allocates nothing of RequestFlow's.
• Resolution is unchanged: every entry into a level resolves its stage, and the bottom level the handler. Stage instances stay the container's, on the lifetime they were registered with.
• NoResultBridge recovers the Task<NoResult> a void stage hands back from its own next call, so a running pass-through level crosses the bridge free.
• docs/lifetimes.md covers registering after BuildServiceProvider: the map gets a plan the provider has no descriptor for, and the dispatch fails on the container's own "No service for type".

Two commits ride along: an ignored Directory.Local.props for per-machine settings, which is where the benchmarks' InternalsVisibleTo lives, and TreatNoTestsAsError, so a test leg whose runner fails to load stops exiting green.

BREAKING CHANGE: IContinuation<TResponse> and IContinuation are gone. A stage takes Continuation<TResponse> or Continuation, both readonly structs over the frozen chain. The calls on next do not change, so migrating a stage means editing one parameter type. Its tests take more: a struct cannot be substituted, so build a real one with Continuation<TResponse>.Over(rest). Invoking a default Continuation throws InvalidOperationException.