perf: cache schedule-shaped simulation routes

[adot-7]

Closes #101

Summary

• precompute immutable schedule timing and cumulative shape geometry during GTFS indexing
• cache active simulation routes by feed generation, Delhi service date, and acceleration configuration
• invalidate on feed reload/error/missing transitions and service/date/config changes without crossing no-service boundaries
• add deterministic cache counters, cache-vs-uncached equivalence tests, and extracted-feed/render benchmarks

Diagnosis and measurements

Against read-only /home/adot/projects/golearning/metroshell/mapdata/extracted on AMD Ryzen 7 8845HS:

• baseline route rebuild: ~9.32 ms/op, 3.68 MB/op, 623 allocs/op (prior temporary benchmark)
• current BenchmarkUncachedExtractedRoutes: 266,258 ns/op, 1.34 MB/op, 143 allocs/op
• current BenchmarkCachedExtractedRoutes: 1,406 ns/op, 16 B/op, 1 alloc/op
• current BenchmarkUncachedExtractedSnapshot: 968,159 ns/op, 1.34 MB/op, 175 allocs/op
• current BenchmarkCachedExtractedSnapshot: 22,968 ns/op, 7,752 B/op, 33 allocs/op
• render benchmarks: normal ~65 us/op; wide ~207 us/op (10 iterations)

Verification

gofmt, go test ./..., go vet ./..., go build ./..., and git diff --check pass.