feat: add opt-in brotli support to the stream compression path

[CodeDredd]

Closes #19

What

Adds brotli to the stream compression path — opt-in, never picked automatically.

app.use(compressionStream())                 // gzip / deflate — unchanged default
app.use(compressionStream({ brotli: true })) // brotli, then gzip, then deflate
app.use(compressionStream('br'))             // always brotli

await useCompressionStream(event, response, { brotli: true })
await useBrotliCompressionStream(event, response)

Brotli stays behind a flag because it is noticeably more CPU-expensive per request than gzip; picking it automatically would silently change the behaviour of every existing compressionStream() user on Accept-Encoding: br, gzip.

How

The native CompressionStream implements the WHATWG CompressionFormat enum — gzip, deflate, deflate-raw — and rejects 'br'. Brotli is therefore streamed through node:zlib, wrapped into a web ReadableWritablePair with Duplex.toWeb(), behind a new createCompressionTransform() seam that both stream call sites (compressStream for the h3 v1 / nitro hook path and compressResponseStream for the v2 middleware) now go through.

BROTLI_OPERATION_FLUSH is load-bearing: with zlib's defaults brotli buffers the whole body until the source closes, which turns a streamed response back into a buffered one. Measured with 5 slowly produced chunks:

variant	emitted output chunks
CompressionStream('gzip')	2
brotli, zlib defaults	1
brotli, BROTLI_OPERATION_FLUSH	6

There is a test asserting the stream stays chunked so this can't silently regress.

No new runtime constraint: src/helper.ts already imported node:zlib statically, so the stream path was never runtime-agnostic. The README now says so explicitly instead of claiming brotli is impossible.

Drive-by fix

useCompressionStream used getAnyCompression and then bailed out on if (compression !== 'br'). For a client sending Accept-Encoding: br, gzip that meant no compression at all — not a gzip fallback. It now uses getStreamCompression, which falls back to gzip. Covered by a test on both the v1 and v2 path.

API

StreamCompression	widened to 'gzip' | 'deflate' | 'br'
StreamCompressionOptions	new — { brotli?: boolean }
CompressionStreamOptions	new — the above plus method?
compressionStream(method | options?)	string argument still works
compressResponseStream(event, value, method?, options?)	new optional 4th argument
useCompressionStream(event, response, options?)	new optional 3rd argument
useBrotliCompressionStream(event, response)	new export

All additive — no breaking changes.

Verification

• pnpm test green on h3 2.0.1-rc.22 (26 passed) and on h3 1.8.1 (10 passed), matching the CI matrix legs
• pnpm lint clean (the 2 remaining warnings are the pre-existing vue/one-component-per-file false positives on createApp in the v1 test)
• pnpm build clean from a wiped dist/, and the dist-bundling regression test still passes — the added node:stream import does not reintroduce the downstream Rollup MISSING_EXPORT problem from fix: read version-specific h3 exports via runtime key #18

New tests: brotli round-trip through the stream on both paths, the default-off fallback, the opt-in pick, forced 'br' without the flag, gzip fallback with brotli enabled but not accepted, and the chunked-output assertion.