feat(media): add AudioMetadataReader.read(ParcelFileDescriptor) overload

[PonceGL]

Stacked on #2821

This branch is cut from chore/p7-fix-audiometautilstest (#2821), not
master directly — it touches the same area (metadata reading) and
shouldn't land without that test class actually running. GitHub can't
target a non-existent branch on the upstream repo, so this PR's diff
includes #2821's commit for now; it'll shrink to just this change once
#2821 merges into master. One of the small independent fixes listed in
#2813.

What

Adds a descriptor-based overload to AudioMetadataReader for reading audio
metadata (title, artist, artwork, lyrics, ...) without needing a filesystem
File — the SAF / downloaded-file path, where a content:// document isn't
guaranteed to have one.

The trap this avoids

The existing read(File) opens its own descriptor and calls
fd.dup().detachFd() for most TagLib calls, but a bare fd.detachFd()
(no dup()) for the last one. That's harmless today only because nothing
touches the descriptor afterwards inside its own use { }. A descriptor
handed in by a caller who keeps using it is a different story:
detachFd() steals the underlying native descriptor, and the caller's own
close() would find it already gone.

Change

• Extracted the TagLib read + property-map-to-field mapping (~90 lines) out
  of read(File) into a shared private buildAudioMetadata(), parameterized
  by a dupFd: () -> Int lambda — avoids duplicating that mapping logic
  across two functions.
• Both overloads now go through dup().detachFd() for every TagLib call,
  including the previously-bare last one. read(File)'s old shortcut is
  gone structurally, not just avoided in the new code.
• read(pfd) never runs the JAudioTagger fallback (it needs a real File,
  which a SAF document may not have) — documented as a known, accepted gap
  on its KDoc. read(File) keeps the full fallback.

Testing

TagLib is a native library; nothing in this repo's JVM tests calls into it
(checked — the one existing TagLib-adjacent test, LyricsRepositoryImplTest,
only covers a pure function around it). This needs a device.

3 new androidTest cases, run on a Pixel_10 AVD, against a 5 KB tagged MP3
(ffmpeg-generated: ID3 title/artist/album + embedded cover):

• read(pfd) returns the same metadata as read(file) for the same file.
• the caller's descriptor is still valid after the call (not detached).
• readArtwork = false skips the embedded cover.

All 3 pass. JVM baseline unaffected (388 tests, 4 pre-existing failures once
#2821 is applied), assembleDebug succeeds.

[PonceGL]

Self-review (code-review, high effort)

Ran a full self-review of this diff before asking for a human one. 6 findings, 3 applied, 2 documented as accepted trade-offs, 1 already covered by this PR's own KDoc:

Applied:

• Both read() overloads now use the existing PerformanceMetrics.time() helper instead of hand-rolled nanoTime/finally timing — that helper already exists and is used elsewhere (MediaItemBuilder.kt); I'd duplicated the pattern across two call sites instead of reusing it.
• read(pfd) gained an optional label param — it was hardcoding "descriptor" in the VERBOSE-gated diagnostic logs, so concurrent reads of different descriptors were indistinguishable in logcat.
• A test assertion force-unwrapped fromFile.artwork!! without an assertNotNull guard first (only fromPfd.artwork had one) — fixed for a clear JUnit failure instead of a raw NPE if it ever regresses.

Accepted, not changed, documented in the commit message:

• The refactor makes read(file)'s artwork read go through dup() even for what used to be a bare detachFd() on the last call — one extra dup()+close() syscall pair per file, on the library-scan hot path. Correctness-neutral, negligible cost; traded for removing the fragile "last call doesn't need dup()" shortcut structurally rather than special-casing it.
• buildAudioMetadata's dupFd and jAudioTaggerFallbackFile are two independent parameters with no compiler-enforced pairing — a sealed MetadataSource type would close that gap, but with exactly two call sites today, it's more machinery than the code warrants yet (GEN-DES-13). Noted for later if a third source shows up.

Already covered: the review flagged that read(pfd)'s hardcoded test assertions depend on TagLib alone parsing the fixture correctly, since there's no JAudioTagger fallback on that path — that's exactly the documented, deliberate scope decision on read(pfd)'s own KDoc (no fallback for SAF-style descriptors), not a new gap.

Re-verified after the fixes: assembleDebug succeeds, JVM baseline unaffected (388 tests, 4 pre-existing failures once #2821 is applied), and all 3 androidTest cases pass on the Pixel_10 AVD.

[PonceGL]

Closing for now — reorganizing how this work is staged. It'll go through our fork first and we'll propose it upstream again, possibly bundled differently, once the larger feature it's part of is further along. Not a rejection, just a process change on our side.