perf: pre-compile middleware chains at build time

[marvinhagemeister]

Pre-compiles middleware chains once at startup instead of constructing them per-request.

Changes

• Replace runMiddlewares(middlewares, ctx, onError) (per-request) with compileMiddlewares(middlewares, onError) (once at startup) which returns a single Middleware<State> function
• Router now stores a single compiled handler per method/route (T | null) instead of arrays (T[])
• The compiled chain passes tail (the original ctx.next) as a parameter to each step, avoiding the infinite recursion bug where a compiled tail reads an already-overwritten ctx.next — safe under concurrent requests
• Lazy middleware resolution is preserved: a lazy middleware is resolved on first request and the resolved function replaces the lazy wrapper in the closure

Benchmarks

Synthetic benchmarks (10 middlewares, deno bench) show the per-request savings are small in isolation — middleware execution and async/await overhead dominate. The main value is architectural: building the chain once is correct and eliminates per-request allocations, which may compound under high concurrency.

Test plan

• Existing middleware tests adapted to compileMiddlewares
• New test: single middleware
• New test: post-processing after next() (response wrapping)
• New test: concurrent requests sharing the same compiled chain
• New test: onError callback invocation
• New test: empty middleware array falls through to ctx.next
• New test: compiled chain calls tail next
• New test: first-registered route wins on duplicate registration
• Router tests adapted to single-item API