Slice stored .conda zip members instead of copying

[andrew]

readCondaMember read each .tar.zst member with io.ReadAll(io.LimitReader(rc, ...)), which grows a buffer by doubling and copies the whole member. Real .conda packages store their members uncompressed (zip.Store), so f.DataOffset() locates the payload inside the outer archive bytes and raw[off : off+CompressedSize64] returns it directly. Bounds are checked against a header that lies about size or offset; non-Store members and out-of-range headers fall through to the existing Open+ReadAll path.

The stored path skips archive/zip's CRC32 check on the member; the inner zstd frame check still catches corruption.

Adds TestCondaDeflatedMemberFallback for the non-Store path and BenchmarkOpenConda (8 MiB incompressible pkg member so the stored .tar.zst is ~8 MiB). Interleaved benchstat, arm64 darwin, n=10, 2s each:

            │   before    │             after              │
            │   sec/op    │   sec/op     vs base           │
OpenConda    4.012m ± 20%   2.452m ± 16%  -38.88% (p=0.000)
            │    B/op     │    B/op      vs base           │
OpenConda    45.98Mi ± 0%   29.69Mi ± 0%  -35.42% (p=0.000)
            │  allocs/op  │  allocs/op   vs base           │
OpenConda      181.0 ± 1%    133.0 ± 0%   -26.52% (p=0.000)

The ~16 MiB B/op saving is roughly 2× the member size (io.ReadAll growth doubles at the last step). Real conda packages run to hundreds of MB per member, so the saving scales with that.