fileserver: reduce allocations in calculateEtag

[jvoisin]

calculateEtag runs for every static file served (and on conditional If-None-Match revalidations). It used a non-presized strings.Builder plus two strconv.FormatInt calls, each allocating an intermediate string and triggering builder growth, which is about six allocations per call.

This commit builds the quoted etag into a fixed 32-byte stack array with strconv.AppendInt instead, so only the final string is heap-allocated. The output format is byte-for-byte identical. Two base-36 int64 values plus the two quote bytes are at most 30 bytes (the worst case is MinInt64, which is 1y2p0ij32e8e8, so 13 digits plus a minus sign, so two times 14, plus the quotes, for a total of 30 characters), so the buffer never spills. And even if it did, which it doesn't append would simply grow it onto the heap without changing the result.

Adds a benchmark, as required by the policy:

CalculateEtag:

• sec/op 347.6n -> 186.2n (-46.42%)
• B/op 112 -> 56 (-50%)
• allocs/op 6 -> 2 (-66.67%)

It removes some CPU work, but more interestingly four of six per-file allocations on a hot path of a very common handler, lowering per-request garbage and GC pressure.

Assistance Disclosure

"No AI was used."