[Mohith26]

Router.lookup runs a regexp for every candidate route (in reverse registration order), and Router.Params runs it again with FindStringSubmatch. Since most services register patterns like /things/:id or /things/:id/*rest, this adds up on every request.

This PR keeps the regexp as the definition of what a pattern means, but skips it when it is not needed:

• newRouteMatcher builds a small matcher at Register time for patterns made of literal text, :params followed by / or the end, and an optional trailing *residual. In that subset there is only one way a path can match, so a single left-to-right scan gives the same answer and the same captures as the regexp. Anything else (/:id.json, /*r/:p, * in the middle, invalid UTF-8 or U+FFFD in the pattern) still uses the regexp, so behaviour is unchanged.
• Serve attached the router, pattern and method with three context.WithValue calls, which is five allocations including the boxed strings. A single routedContext now answers all three keys and delegates everything else (values, Done, Err, Deadline) to the parent, so cancellation still propagates.

Testing:

• TestRouteMatcherAgreesWithRegexp runs ~1,300 pattern/path pairs (literals, params, residuals, newlines, unicode, invalid UTF-8) and checks the matcher agrees with the compiled regexp on both the match and the captured params.
• FuzzRouteMatcherAgreesWithRegexp does the same with fuzzed inputs; I ran it for 150s (about 11.7M executions) with no differences.
• TestRouterMixedMatchersPrecedence interleaves fast-path and regexp routes to check last-registered-wins still holds, and TestRoutedContext covers the context values, shadowing, cancellation and the nil-parent panic.
• go vet ./..., go test ./... and go test -race on the router tests pass.

Benchmarks (go test -run '^$' -bench BenchmarkRouter -benchtime 200ms, old and new runs interleaved, benchstat medians, Apple Silicon arm64). I also added BenchmarkRouterManyRoutes, a 40-route table shaped like a typical service. The timings are from a busy shared machine so they are noisy; the allocation counts are exact.

benchmark	before	after	allocs
Lookup GET /index.html	368 ns	78 ns	1 -> 1
Lookup GET /1/2 (two params)	320 ns	172 ns	4 -> 2
Lookup GET /404	231 ns	100 ns	1 -> 1
Serve GET /index.html	540 ns	185 ns	6 -> 2
Serve GET /1/2	277 ns	147 ns	6 -> 2
ManyRoutes Serve, last-registered route	397 ns	152 ns	6 -> 2
ManyRoutes Params, first-registered route	825 ns	349 ns	3 -> 2
ManyRoutes Params, last-registered route	557 ns	152 ns	4 -> 2

Across all router benchmarks the geomean is -38% time and -46% allocations. The not-found path is unchanged since its cost is building the terror.

I used an AI assistant while writing this; I reviewed the change and ran everything above locally.