refactor(rlp): tighten encode/decode, unify length-prefix logic, lift docs

[tcoratger]

Summary

• Collapse encode bytes + encode list + encode length into one shared length-prefix helper parameterised by base (0x80 for strings, 0xC0 for lists). Same idea on the decode side: four near-identical short/long string/list branches fold into a single length helper that returns the payload range.
• Remove a defensive bounds check in the decode dispatcher that turns out to be unreachable from either call path (top-level rejects empty input, recursive call is gated by the parent list's payload end which is bounded by the buffer length).
• Rewrite every docstring and inline comment to match the bitfields module style: one sentence per line, no backticks, role-based prose (no function/variable names in doc text), Phase N labels with concrete hex walkthroughs at each algorithm step.
• Inline the magic numbers (0x80, 0xB7, 0xC0, 0xF7, 55) since they are the lingua franca of the RLP spec; previously named module constants moved into the test file as locals.
• Add tests for three previously uncovered error paths:
  • leading-zero length bytes rejected for long strings
  • leading-zero length bytes rejected for long lists
  • inner item overshooting parent list boundary

Result

• Source: 302 -> 333 lines (executable code shrinks from ~130 to ~75 lines; the rest is now richer documentation).
• Coverage: RLP module now at 100% line and branch coverage (77/77, 38/38).

Test plan

• uv run pytest tests/lean_spec/types/test_rlp.py — 352 passed
• uv run pytest tests/lean_spec/test_cli.py tests/lean_spec/subspecs/networking/enr/test_enr.py — consumers still green
• just check — ruff, format, ty, codespell, mdformat all pass
• 100% line + branch coverage on src/lean_spec/types/rlp.py