Use pointer-based pixel storage in memory efficient replay buffer

[rojas70]

Motivation

• Reduce memory duplication from storing full pixel stacks per transition by storing only integer pointers in the replay dataset and keeping pixel frames in a shared ring buffer.
• Keep existing API for sampling stacked observations while significantly lowering memory pressure for large capacities and long stacks.

Description

• Allocate per-pixel-key ring buffers (self._pixel_buffers) and expose observation pixel entries as int32 pointer Box spaces instead of full images.
• Add helper methods _insert_pixel_frame and _advance_pixel_pointer to write frames into the ring buffer and advance the frame pointer.
• Update insert() to write incoming pixel frames into the ring buffer and store pointer indices in observations/next_observations instead of copying images into each transition.
• Update sample() to reconstruct stacked observation tensors by resolving pointer histories from the dataset and indexing into the pixel ring buffer (replacing the previous sliding_window_view over stored images).

Testing

• Successfully type-checked/compiled the modified file with python -m py_compile serl_launcher/serl_launcher/data/memory_efficient_replay_buffer.py.
• Ran a small runtime validation script that constructs the buffer, inserts random transitions, and samples a batch to verify output shapes and that dataset pixel entries are int32 pointers; the script completed and returned expected shapes and pointer dtype.

---

Codex Task

[rojas70]

also done with code/chatGPT have not tested. but @ryanvanderstelt wanted to throw this idea out into the datastore...