docs: demonstrate W_pos migration

[lntutor]

Description

Adds a runnable GPT-2 migration example for the HookedTransformer.W_pos and HookedTransformer.W_E_pos mappings. The example:

• reads bridge.pos_embed.W_pos
• constructs W_E_pos from bridge.W_E and W_pos
• verifies the expected tensor shapes
• demonstrates equality with the legacy properties under matching compatibility-mode weight processing

The accompanying explanation also points readers back to the existing numerical-matching guidance and mentions the raw/no-processing alternative.

Fixes #1560

Type of change

• This change requires a documentation update

Verification

• make format
• Executed the documented GPT-2 snippet; both shape assertions and torch.testing.assert_close comparisons pass
• uv run build-docs (build succeeded, 4 warnings; all four are pre-existing optional-import/duplicate-object warnings)
• git diff --check

Checklist

• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

Not applicable: code comments and new automated tests, because this is a documentation-only runnable example.

[lntutor]

The full-coverage job's only failure is an unrelated flaky numerical-tolerance assertion in tests/unit/test_direct_path_patching.py after 6,597 tests passed. This PR only changes docs/source/content/migrating_to_v3.md. I don't have permission to rerun the workflow; could a maintainer please rerun the failed job?

[lntutor]

The failing Full Code Coverage job appears unrelated to this docs-only change. It completed 6,597 tests and failed only tests/unit/test_direct_path_patching.py::TestCorrectness::test_correctness_against_actual_ln_forward on a numerical tolerance assertion.

I reproduced that exact test on the current PR head 086e4c1c with the CI Python version:

uv run --python 3.12 pytest tests/unit/test_direct_path_patching.py::TestCorrectness::test_correctness_against_actual_ln_forward -q
1 passed

It also passes under Python 3.13 locally. The PR changes only docs/source/content/migrating_to_v3.md; the failing module is untouched. The test's module-scoped random model is constructed without resetting its seed immediately before construction, so xdist scheduling can change the generated weights even though the session seed is fixed.

I do not have upstream workflow rerun permission. Could a maintainer rerun the failed Full Code Coverage job for this exact head?

[jlarson4]

@lntutor I am rerunning it now sorry just getting to your comments! Lots of contributions to go through this morning.

This specific CI test has been producing sporadic failures lately, it is a known issue that I have not gotten around to addressing. Thank you for the specific details about the failing test, I have noted this PR down in my notes for this problem and will come back here to reference it when I get around to fixing it!

[jlarson4]

Flaky test is still failing, so I am pushing a fix to dev-4.x. I will make sure to pul it in so this can pass CI as soon as its ready.

[jlarson4]

This looks great, thanks @lntutor! Just one small comment