transport/http: direct header insertion on HTTP/2 outbound

[bananacocodrilo]

RELEASE NOTES:

On HTTP/2, header field names are always lowercased on the wire (RFC 7540 §8.1.2), so the canonicalization that net/http performs via textproto.CanonicalMIMEHeaderKey on every http.Header.Set/Add call is redundant and only adds CPU and allocations.

This PR introduces a fast path on the HTTP/2 outbound that writes directly into the underlying http.Header map, bypassing canonicalization in three hot spots:

• application header conversion — headerMapper.ToHTTPHeadersPreserveCase
• core RPC header writes — withCoreHeaders (gated by o.useHTTP2)
• OpenTracing header injection — directHTTPHeadersCarrier

The HTTP/1.1 path is unchanged and continues to canonicalize header names, preserving backward compatibility with peers that perform case-sensitive header matching.

Why it's safe

Path	Behavior	Reason
HTTP/2 outbound	New: keys written as stored (no canonicalization)	HTTP/2 wire format mandates lowercase header names; map casing is irrelevant on the wire
HTTP/1.1 outbound	Unchanged	Still uses http.Header.Set/Add and opentracing.HTTPHeadersCarrier
AddHeader (setup-time)	Unchanged	Still uses http.Header.Add; not a hot path
Core RPC header constants (CallerHeader, ServiceHeader, …)	Already canonical (Rpc-Caller, Rpc-Service, …)	Stored casing matches what canonicalization would produce

Wire-format implications

• HTTP/2: header names appear lowercase on the wire (HTTP/2 requirement) regardless of this change. No observable difference for peers.
• HTTP/1.1: no change.
• Tracing headers on HTTP/2 are now stored with whatever casing the tracer supplies (typically lowercase, e.g. uber-trace-id) instead of the canonicalized form (Uber-Trace-Id). On the wire it's still lowercase per HTTP/2 spec.

Performance

Benchmarks on Apple M2 Max, count=6, full createRequest + withCoreHeaders path. Median of 6 runs.

Scenario	http1_canonical	http2_direct	Δ time	Δ allocs
4 headers, mixed case (typical)	2440 ns / 37 allocs	1819 ns / 32 allocs	−25%	−14%
24 headers, all lowercase	5716 ns / 96 allocs	3275 ns / 71 allocs	−43%	−26%
24 headers, mixed case	5677 ns / 96 allocs	3478 ns / 71 allocs	−39%	−26%

Bytes allocated drop 4–11% across scenarios.