http3: encode response headers against the QPACK static table

[MDA2AV]

What

Qpack.EncodeResponseFields consulted the static table for :status alone. Every other field went out as Literal With Literal Name, spelling out a name the peer already holds by index — content-type cost 12 bytes on the wire on every response, forever.

Encoding is now three tiers, cheapest first:

case	encoding	cost
name and value in the table	Indexed Field Line (0xC0)	1 byte
name in the table	Literal With Name Reference (0x50)	index + literal value
neither	Literal With Literal Name (0x20)	unchanged

Names match case-insensitively (so a caller holding conventional capitalisation resolves, and the name is never written); values match exactly, since a field value is case-sensitive.

The lookup has to be cheap, or it costs more than it saves

Worth calling out, because the first version was a regression. A plain linear scan over the ~50 distinct static names, per header, cost 6% of throughput — more than the bytes it saved. Bucketing by name length narrows it to one or two comparisons, and the saving flips to a gain:

	run 1	run 2	run 3
literal names (before)	3,087,560	3,123,721	3,171,774
static table, linear scan	2,771,482	2,941,583	2,993,596
static table, bucketed	3,291,610	3,316,463	3,367,788

h3x -t 8 --connections 8 -m 100, buffered playground. Median 3,123,721 → 3,316,463 req/s, +6%.

One thing left alone deliberately

Http3Response.Text sends content-type: text/plain; charset=utf-8 — with a space. Static entry 54 is text/plain;charset=utf-8 without one, so it misses the one-byte indexed encoding and takes the name reference instead. Dropping the space would make it an exact hit, but that changes what goes on the wire, so it is not in this PR.

Tests

Unit 31 / E2E 51 / Chaos 47, all passing. Notably http3: pure-C# parser GET vs nghttp3 client still passes, which is the check that matters — our encoder's output decoded by an independent QPACK implementation.

New unit tests reach Qpack via InternalsVisibleTo rather than by widening the public surface. They verify by decoding what the encoder produces: an index that is off by one still yields a well-formed field section, so only a round trip catches it naming the wrong header. Covered: exact matches, known-name-unknown-value, capitalised names, unknown names lowercased on the way out, and case-sensitive value matching.

Not addressed

ioxide.http3 has no QPACK dynamic table at all — SETTINGS advertise capacity 0 and blocked streams 0, and the peer's SETTINGS values are walked for framing and discarded. That is unchanged here. Measured across clients, only Chrome advertises any capacity and nothing inserts inbound, so a dynamic table would be a browser-only win; this change benefits every client instead.