feat: streaming encoder (io::Write wrapper)

[polaz]

Summary

The current encoder requires all input data upfront via Read trait. There is no streaming encoder that implements io::Write — callers cannot incrementally feed data and get compressed output. This is required for network streams, large files, and pipe-based workflows.

Current state

• FrameCompressor::compress() reads all data in a loop until EOF
• No Write-based wrapper analogous to StreamingDecoder for decompression
• CLI buffers entire input before compression

C reference implementation

Streaming compression API

ZSTD_compressStream2(cctx, &outBuf, &inBuf, endDirective)
// endDirective: ZSTD_e_continue, ZSTD_e_flush, ZSTD_e_end

• ZSTD_e_continue — accumulate input, compress when block is full
• ZSTD_e_flush — flush current data (partial block)
• ZSTD_e_end — finalize frame with last block flag + optional checksum

Key design patterns

• Input buffer accumulation until block size reached (128KB max)
• Partial flush support for latency-sensitive use cases
• Frame finalization separate from data feeding

What needs to be implemented

1. StreamingEncoder<W: Write> — wraps a Write drain
2. impl Write for StreamingEncoder — accumulates input, emits compressed blocks when buffer full
3. flush() method — flush partial block (non-last)
4. finish() method — write last block + checksum, return inner writer
5. Frame header — written on first write() call
6. Block buffering — accumulate up to 128KB before compressing block

Acceptance criteria

• StreamingEncoder implements io::Write
• Can compress data incrementally (multiple write() calls)
• finish() properly closes the frame
• Interop: C zstd decompresses streaming-encoded output
• Roundtrip: StreamingEncoder → StreamingDecoder works
• CLI uses streaming encoder for pipe support

Time estimate

2d