Add PagedDecoder trait for multi-page formats (TIFF, PDF, DICOM)

## Summary

Multi-page formats like TIFF, PDF, and DICOM contain independent images (pages) that differ in dimensions, color space, bit depth, ICC profile, resolution, and orientation. The existing `FullFrameDecoder` trait is animation-oriented (same canvas, temporal ordering, compositing) and doesn't fit.

## Proposed API

```rust
trait PagedDecoder {
type Error;

/// Metadata for the page at `index`, without decoding pixels.
/// Returns `None` if `index` is past the last page.
fn page_info(&mut self, index: u32) -> Result, At>;

/// Decode the page at `index` to pixels.
/// Returns `None` if `index` is past the last page.
fn decode_page(&mut self, index: u32) -> Result, At>;
}
```

## Design decisions

- **No `page_count()`**: TIFF stores pages as a linked list of IFDs — the count isn't in the header and requires walking the entire chain. Instead, callers discover the last page by getting `None` back. This avoids forcing an upfront scan.
- **Reuse `ImageInfo`**: Each page is an independent image. `ImageInfo` already carries dimensions, format, alpha, orientation, source color (ICC, CICP, bit depth), and embedded metadata. Unused fields (animation, gain map) default to harmless values. No new metadata type needed.
- **`page_info()` is cheap**: Reads the IFD/page header (tag parsing only, no pixel decompression). Useful for building a page list UI or selecting which pages to decode.
- **Random access by index**: Pages are 0-indexed in file order. The underlying `tiff` crate caches IFD offsets as they're visited, so repeated access to the same page is O(1) after first visit.
- **Fits existing trait hierarchy**: `PagedDecoder` is another execution mode from `DecodeJob`, parallel to `Decode`, `StreamingDecode`, and `FullFrameDecoder`.

## Per-page metadata (TIFF)

Every TIFF tag is per-IFD. Each page can independently vary:
- Dimensions, bit depth, sample format
- Color space (PhotometricInterpretation), ICC profile
- Compression method
- Resolution/DPI, orientation
- Extra samples / alpha
- Planar vs chunky layout

## Why not `FullFrameDecoder`?

| | Animation (GIF/APNG/WebP) | Pages (TIFF/PDF) |
|---|---|---|
| Dimensions | Same canvas | Different per page |
| Temporal ordering | Yes (duration_ms) | No |
| Compositing/disposal | Yes | No |
| Random access | Nice-to-have | Essential |
| Per-image metadata | Shared | Independent |

Start by locating DecodeJob and the existing Decode, StreamingDecode, and FullFrameDecoder traits, then read ImageInfo, DecodeOutput, and At to understand how a parallel execution mode fits. Done means defining the PagedDecoder API with per-page metadata and pixel decoding, including None for indexes past the last page and support for independent page properties.

Written by the indexing model from the issue text.