Snapshot initial observation padding before source buffers are reused

[sylvesterkaczmarek]

Summary

Fixes #559.

Copy the initial observation before using it for initial-value buffer padding. Those padding entries previously retained the caller's array, so later mutation or reuse of a MuJoCo state view rewrote the apparent initial observations. Regular inserted observations already copy their input.

Keep a single private snapshot for the padding entries. Buffer reads remain independent copies, and zero padding, arrival ordering, later samples and source arrays are unchanged. This also makes delayed scalar/list padding use the same ndarray representation as regular entries. No dynamics, rewards, dependencies or workflows change.

Validation

Added 21 cases covering source reuse, immediate/delayed observations, scalar/list/strided/Fortran inputs, several buffer sizes and numeric dtypes, returned-array independence, delayed arrival controls and zero padding. Two tests run the actual Updater with a native MuJoCo sliding body whose observable returns its live qpos array, with and without mean aggregation.

• New cases against unchanged main: 19 failed, 2 passed.
• Selected observation, variation, RL and utility suites: 189 passed, 6 camera-related cases deselected, on both Python 3.12.11 and 3.11.16.
• Both independent observation fixes from this round in a separate worktree: 221 passed, 6 deselected, on both Python versions.
• Ruff E9/F on both changed files, Pyink on the added tests and changed source lines, compilation and git diff --check: passed.

Based on main at a04e3e4cf56c12117d2294bb090f9acec21e5c67. Tested on macOS with MuJoCo 3.14.0, NumPy 2.3.2 and SciPy 1.16.1; bindings were generated using the repository's build_mjbindings command. Rendering was disabled. A broader run failed the same three camera-rendering tests on unchanged main and both branches; the CPU selection above excludes camera-related cases only, without modifying repository test settings. The existing axis-angle warning is unchanged. The complete repository suite, rendering, training runs, package builds and other operating systems were not tested. Nested objects in object-dtype arrays are not deep-copied, matching ordinary insertion semantics.

Reference: NumPy array copy semantics.

Upstream checks

At submission the CLA passed, the GitHub Actions Scan was queued, and the organization security workflow reported action_required: https://github.com/google-deepmind/dm_control/actions/runs/36484650446. The test counts above are local results, not completed upstream CI.

[sylvesterkaczmarek]

@saran-t This is current with main, has no failing or pending public checks, and touches the delayed observation-buffer behavior you originally introduced. It snapshots the initial padding value so later reuse/mutation of the source buffer cannot alter already-buffered history, with focused regression coverage. Could you review the current head when convenient?