opj_j2k: fix invalid tile part index error when decoding tiles with interleaved tile-parts

[mbklein]

Fixes #1558

Problem

opj_get_decoded_tile() failed with "Invalid tile part index for tile number N. Got X, expected 0" on codestreams whose tile-parts are interleaved across tiles (all tiles' first tile-part, then all tiles' second tile-part, etc.) and that only declare the total tile-part count (TNsot) in each tile's last tile-part.

Cause

When the tile is decoded, the prior tile decode stops as soon as its own last tile-part is consumed, which can be reached before the target tile's final tile-part (the one containing TNsot). The target tile's nb_tps is never set and its tile-part start positions are never recorded, so opj_j2k_decode_one_tile() takes the "index not built" fallback. That fallback seeks to m_last_sot_read_pos, which for an interleaved layout points past the tile's first tile-part, causing the decoder to encounter a non-zero TPsot first and abort.

This only occurs when multiple tiles are decoded from the same codec (e.g. j2k_random_tile_access). The first tile always succeeds because the index hasn't been built yet.

Fix

• Seek to the start of the codestream (main_head_end + 2) instead, so the target tile's tile-parts are read in order. The faster index-based path used by ordered codestreams is unchanged.
• Add a self-contained non-regression test (rta_interleaved_tile_parts): it encodes a small multi-tile image, rewrites it into the interleaved layout, then decodes every tile and checks the samples against a full-image decode. The test fails before this fix and passes after it, with no external test data dependency.

[mbklein]

That one failing build action is due to a download hiccup, but it doesn't seem to be something I can retry myself.