docs: warn against using Gymnasium FrameStackObservation for image observations

[midhunxavier]

Description

Adds a warning to the custom-environment documentation explaining that Gymnasium's FrameStackObservation wrapper adds a new leading dimension to image observations (e.g. an image of shape (3, 64, 64) becomes (2, 3, 64, 64) when stacking 2 frames). SB3's is_image_space check requires exactly 3 dimensions, so the stacked space is treated as a batch of images rather than an image. As a result, a FlattenExtractor is silently used instead of a CNN.

The note recommends using SB3's VecFrameStack instead (which stacks along the channel dimension and keeps the observation recognized as an image) and links to the issue. A changelog entry is added accordingly.

This follows @araffin's guidance in the issue to keep the is_image_space check as-is and document the pitfall instead of changing the behavior.

Motivation and Context

closes #2090

Users stacking image frames with Gymnasium's FrameStackObservation get a silent fallback to FlattenExtractor (no CNN), with no error, which is surprising and hard to debug. Documenting the recommended VecFrameStack alternative prevents this pitfall.

• I have raised an issue to propose this change (required for new features and bug fixes) — issue [Bug]: is_image_space works poorly with Gymnasium's FrameStackObservation #2090, triaged by a maintainer (@araffin), who proposed this documentation approach.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation (update in the documentation)

Checklist

• I've read the CONTRIBUTION guide (required)
• I have updated the changelog accordingly (docs/misc/changelog.md) (required).
• My change requires a change to the documentation.
• I have updated the tests accordingly (required for a bug fix or a new feature). — N/A: documentation-only change, no code modified.
• I have updated the documentation accordingly.
• I have opened an associated PR on the SB3-Contrib repository (if necessary) — N/A.
• I have opened an associated PR on the RL-Zoo3 repository (if necessary) — N/A.
• I have reformatted the code using make format (required) — N/A: no Python code changed (docs only).
• I have checked the codestyle using make check-codestyle and make lint (required) — N/A: no Python code changed (docs only).
• I have ensured make pytest and make type both pass. (required) — N/A: no Python code changed (docs only).
• I have checked that the documentation builds using make doc (required) — build succeeded, no warnings/errors; the warning admonition renders on the custom-env page.

---

LLM/code-assistant disclosure (per CONTRIBUTING.md): this documentation change was prepared with the assistance of Claude, an LLM code assistant. The approach — documenting the pitfall and recommending VecFrameStack rather than altering is_image_space — was proposed by maintainer @araffin in #2090. I have personally reviewed and verified the change.