🩹 Confine the zero-count early-out to the half-float path

[MarijnS95]

read_compressed_data() returns before touching the reader whenever count is zero, which desyncs the stream for every grid that is not stored as half floats. OpenVDB only skips the read in HalfReader</*IsReal=*/true, T>::read(), which bails on count < 1 before consuming anything; every other read goes through readData() into bloscFromStream()/unzipFromStream(), and those always consume the block's i64 size header plus its payload regardless of the count. Returning early leaves the reader parked mid-block, so every subsequent read comes back shifted and eventually trips the assert_eq!(-compressed_count as usize, count) in that same function, reading a num_compressed_bytes of 0 where a real count was expected.

#63 reports that same assertion, from before the early-out existed, and was never closed — #73 landed the blanket early-out and closed #33 instead. Since the file in #63 is one of the Embergen exports it is plausibly already covered by the half-float branch, so I am not claiming this closes it; what this fixes is the same assert firing for grids that are not half floats. That split also explains the loose end noted in #73, where Embergen's f16 exports came out fixed while the Houdini f32 exports stayed broken: f16 grids take the branch where the early-out genuinely belongs, and f32 grids take the path where OpenVDB always consumes the block. This moves the early-out into the half-float branch so it mirrors HalfReader, and leaves a note in read_compressed_data() explaining why a blanket one cannot come back.

For evidence I read Breda's 30 MB Nubis cloud-noise VDB, four ACTIVE_MASK | BLOSC f32 grids that OpenVDB writes with zero-count blocks. On main it panics with left: 0, right: 204. With this change all four grids parse, and an order-sensitive checksum over every voxel matches what the last release before #73 (vdb-rs 0.6.0 from crates.io) produces, so the values are right and not merely non-crashing:

lf_curly_worley: voxels=2091020 sum=258537.719319 checksum=0xbb49a9094874c44e
hf_curly_worley: voxels=2097152 sum=283849.192542 checksum=0xc71c704c6a1de110
lf_alligator:    voxels=2097152 sum=815082.756296 checksum=0xf0ffa9d0778efb5f
hf_alligator:    voxels=2097152 sum=831611.971946 checksum=0xe946b3983db57769

The half-float path is untouched by construction, and OpenVDB's smoke.vdb and fire.vdb (whose density grids are half) give byte-identical checksums before and after. I could not find a grid that is both half-float and BLOSC to exercise that combination directly — the ASWF sample models predate BLOSC, and the Embergen files from #73 are not something I have — so that pairing rests on the early-out now sitting exactly where OpenVDB puts it rather than on a measurement.