Optional 16‑bit sample logging (true 16‑bit PNGs, TB previews stay 8‑bit) for unconditional example

[turian]

Related to #12596

Purpose

Allow users to optionally export 16-bit evaluation samples (PNG) from the unconditional training example while keeping TensorBoard previews 8-bit and avoiding silent dtype conversions.

Changes

• Adds helpers:
  • _prepare_sample_images: clamps/scales NHWC float [0,1] to uint8/uint16 arrays; also builds an 8-bit preview for TensorBoard.
  • _log_sample_images: logs previews to TensorBoard and uploads true 16-bit PNGs to W&B via file paths.
• Introduces --image_bit_depth with choices {8, 16} (default 8). Explicit guards raise ValueError for unsupported depths/dtypes.
• W&B: replaces in-memory buffer usage with temporary file paths, so the uploaded artifact is the exact 16-bit PNG. Notes:
  • Pillow is used to encode 16-bit PNGs. BytesIO is avoided because wandb.Image does not reliably accept it.
  • For RGB(A) 16-bit, uses raw mode "RGB;16B"/"RGBA;16B" and byteswap to satisfy Pillow’s big-endian expectation.
• Aligns image logging steps to global_step (matches scalar metrics).
• Docs: README documents the flag and clarifies preview behavior (TensorBoard stays 8-bit; W&B previews are typically 8-bit, but uploaded files keep 16-bit).

Behavior/UX

• Default behavior unchanged (8-bit).
• When --image_bit_depth 16 is set:
  • TensorBoard shows 8-bit previews for usability.
  • W&B receives the true 16-bit PNG files; previews may look 8-bit in the UI, but the stored files are 16-bit and can be downloaded.
• Explicit dtype/shape checks fail fast if inputs are not uint8/uint16 or are in an unsupported layout.

Limitations (intentional to keep scope small)

• 32-bit TIFFs are not included (would require broader encoding and logger plumbing).
• Pillow is required at runtime for W&B 16-bit encoding. We keep the import scoped and do not add a new dependency to the example’s requirements to avoid expanding the PR; torchvision typically brings Pillow in.

Notes for reviewers

• No change to training semantics; this only touches evaluation/sample logging.
• Helper functions localize the conversion and logging behavior for future extensions (additional bit depths can be added by extending the helpers and CLI choices).
• Endianness is handled explicitly for 16-bit RGB(A) via ";16B" and byteswap.
• Temporary file paths are unlinked after logging.

Follow-ups (out of scope here)

• Optional: support 32-bit float/TIFF and a generic bit-depth pipeline.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

• Training examples: @sayakpaul

[turian]

Closing for now, I need to test this more.