fix: time all URL parsers on the same inputs

[lemire]

Summary

The top-100k URL benchmarks were not comparable. Each BenchmarkTop100* function break'd on the first parse error, and the three libraries reject different inputs:

• goada and whatwg-url fail first at index 5063 (plain text, no scheme)
• net/url continues until index 7828

So goada/whatwg-url were timed on 5,063 URLs while net/url was timed on 7,828 — and none of them ran the full dataset. The README then divided ns/op by ~100k, which made every library look ~20× faster than it was.

Changes

• Parse every URL in the shared list, including ones that fail
• Skip empty lines when loading top100.txt (6 lines; not URLs)
• Keep parse results in sinks so the compiler cannot dead-code the calls
• Fail the benchmark if the dataset failed to load

Remeasured numbers

Median of 7 runs (-benchtime=3s -count=7) on Apple M4 Max, 100,025 URLs:

Library	ns/op (full dataset)	ns/op per URL	WHATWG
goada	58,533,792	585.2	yes
Go net/url	31,780,421	317.7	no
whatwg-url	322,011,608	3219.3	yes

Reproduce with:

go test -bench BenchmarkTop100 -run -