docs: TE accuracy — flag zero-crossing rows; remove misleading "Known limitation"

[MinhMPA]

Summary

Documentation-only fix (no clax module code changed). Reframes the README's unlensed-C_l^TE accuracy presentation around the two ΛCDM TE zero crossings (ℓ≈52, ℓ≈400) and removes a "Known limitation" bullet that mischaracterizes a metric artifact as a physics shortcoming.

Why

The current README accuracy table reports relative error (clax − CLASS) / CLASS at every multipole, including ℓ values where C_l^TE crosses zero. Near a zero crossing the denominator goes to ~0, so the relative number can blow up to ~0.5–1% even when the absolute residual matches neighboring ℓ where the same metric is well-defined. The current "Known limitations" entry — "TE zero crossings: Large relative errors near l=52 and l=400 where C_l^TE crosses zero" — frames this as a real shortcoming, when it is purely a presentation issue. The underlying C_l^TE matches CLASS as well as TT/EE do.

What changes

• README accuracy table: added a † marker on the three rows clearly inside the first zero-crossing region (ℓ=20, 30, 50) and a footnote that:
  • states the relative-error metric is ill-defined there,
  • points at the Hu & White (1997) correlation criterion |C_l^TE| / √(C_l^TT · C_l^EE) < 0.02,
  • notes that the existing lensed-TE accuracy test (tests/test_lensing.py:156) already skips zero-crossing ℓ via this criterion, and
  • notes that a Gaussian likelihood weights these modes by 1/Var(C_l^TE) → 0 automatically — so the metric artifact does not affect HMC inference.
• README "Known limitations": removed the "TE zero crossings" bullet. The remaining items in that section (speed, TT ℓ=400-800, TT ℓ>1200, EE ℓ=20-30, primordial BB) are all genuine outstanding limitations.
• CHANGELOG: entry documenting the rationale.

No test changes needed: tests/test_harmonic.py::TestClTE only probes ℓ=100 and ℓ=200, neither near a zero crossing.

Scope clarification

The ℓ=1000 TE entry (+1.7%) is not a zero-crossing artifact — it is a real residual driven by k-grid under-resolution at high ℓ (same root cause as the existing "TT ℓ>1200" known limitation). That one will be addressed separately by a hybrid linear/log k-grid PR; this PR does not touch it.

Independence from other open PRs

• No file overlap with refactor: collapse 6 compute_cl_pp_* into single compute_cl_pp(... nonlinear=...) — redo of #11 with bug fixes #14, refactor: vmap-safe transfer + nonlinear; drop _halofit_modulator inline duplication #15, or fix(compute_pk): integrate to full bg.conformal_age + recover dtau_end gradient via Taylor correction #16.
• Doesn't touch clax/ modules, tests, or any code path PR feat: clax-pt — differentiable one-loop EFTofLSS galaxy power spectra in JAX #9 (clax-pt draft) edits.
• Should rebase cleanly behind any of those.

Test plan

• No code changed; existing test suite is unaffected.
• Existing tests/test_harmonic.py::TestClTE (ℓ=100, 200) probe ℓ values that are not near zero crossings, so the documentation change is consistent with what the tests already exercise.
• Existing tests/test_lensing.py::TestLensCls::test_lensed_te_accuracy already implements the correlation-coefficient skip rule that the README footnote now references.

[MinhMPA]

PR #17, #18 and #19 aim to address the "Known limitations" in the current README so that we can clean up that section. These should be independent fixes so no merge order nor conflict is expected.