[BugFix] Harden memmap path handling

[peterdsharpe]

Summary

• Apply robust filesystem encoding consistently to nested TensorDict directories, dynamically created subtrees, and subpath resolution.
• Retain legacy-layout fallback for safe single-component names.
• Reserve new memory-map targets atomically, reject symbolic-link targets, and preserve the existing no-file behavior for zero-element tensors.
• Clarify filesystem ownership and overwrite expectations in the storage documentation.

Rationale

Nested collections previously did not use the same filesystem-key handling as tensor leaves. This aligns save and load behavior across both cases while retaining the explicit robust_key=False compatibility path and existing overwrite workflows.

Validation

• 1760 passed, 54 skipped across memmap-focused constructor, TensorDict, and round-trip tests
• Focused robust-key, legacy-layout, symbolic-link, and key-codec checks
• Repository-pinned ufmt, flake8, and pydocstyle checks
• RST-title and diff validation

[vmoens]

LGTM thanks!
Can you have a look at 0211cce before I merge?

[peterdsharpe]

LGTM thanks! Can you have a look at 0211cce before I merge?

Good catch! Yep, LGTM!