docs: add release notes for 1.1.0 with known regressions and workarounds

[andygrove]

Which issue does this PR close?

None. It comes out of the 1.1.0 regression audit (#6399), and the regressions it lists are tracked in #6402.

Rationale for this change

Comet publishes a change log for each release, but no release notes. The upgrade guide already says that correctness fixes "appear in the release notes", and the audit of 1.1.0 found regressions that users should know about before they upgrade, along with the settings that avoid them.

What changes are included in this PR?

• A new top-level Release Notes section in docs/source/release-notes/, with an index page and a page for 1.1.0. It sits before the change log in the site navigation.
• The 1.1.0 page opens with how many 1.0.0 bugs the release fixes, then lists the one known regression that 1.1.0 still has, A native final aggregate that has spilled can fail the task during its replay #6254, with a link to its issue and the settings that avoid it. A last entry covers fix: revert unsafe partial aggregates after final fallback #5421, which moves some partial aggregates back to Spark when the Comet shuffle manager isn't installed. That one is a deliberate trade for correctness rather than a regression, but it can make those jobs slower.
• The upgrade guide's 1.1.0 section links to the release notes, and gains a subsection on native Iceberg reads with IRSA. Since fix: Native S3 scan on EKS/IRSA turns a transient STS throttle into a hard 403 storm #6025, they no longer fall back to the node role when the web-identity call fails. The subsection names the per-catalog opt-out, and the setting to use for tables loaded by path.

1.1.0 will be released from an rc2 cut from branch-1.1, which has the fixes for #6424 and #6464 (#6449), #6423 (#6489), #6426 (#6486), #5507 (#6487), #6505 (#6511), #6334 (#6491), #6466 (#6488), #6506 (#6540), #6425 (#6554), #6504 and #6546 (#6560) and #5701 (#6561), so those aren't listed. #6546 turned up after the audit, while fixing #6504. If #6544, the fix for #6254, reaches branch-1.1 before rc2, that entry should come out too.

The bug counts come from the audit's data. rc1 has 131 fix: pull requests. 91 of them fix bugs that shipped in 1.0.0, and about 40 of those 91 fixed results that differed from Spark.

How are these changes tested?

Each workaround was run against its regression's reproducer on a 1.1.0-rc1 build. Without the setting the reproducer still differs from Spark, and with it the result matches Spark.

Prettier passes on the changed files. I couldn't build the Sphinx site locally, and the docs job only runs on pushes to main, so the new section hasn't been rendered yet.

[andygrove]

Closing this. Every regression it listed is now fixed on branch-1.1 except #6254, and #6544, its fix, is approved and labelled for backport, so the 1.1.0 page would be empty or close to it. That isn't worth a new top-level docs section. The part users still need, the upgrade-guide note that native Iceberg reads on EKS no longer fall back from IRSA (#6025), is now #6603, and it will be backported to branch-1.1. #6402 keeps the full list of regressions and their fixes.

Two things from here belong in the 1.1.0 release blog post instead:

• The headline from the audit (Audit the PRs in 1.1.0 for regressions since 1.0.0 #6399): 1.1.0 fixes about 90 bugs that shipped in 1.0.0 (91 of rc1's 131 fix: pull requests), and about 40 of those returned results that differed from Spark.
• The fix: revert unsafe partial aggregates after final fallback #5421 change: when the final aggregate runs in Spark, which happens for every aggregate if the Comet shuffle manager isn't installed, the partial avg, decimal sum, stddev, variance, corr, first, last and a few others now run in Spark too. In 1.0.0 they ran natively, and avg could return NULL in that plan (Unsafe native partial aggregates survive child-triggered final fallback #5419). Installing the Comet shuffle manager keeps them native.

If #6544 doesn't make rc2, the #6254 workaround can go in the post too: more off-heap memory, or spark.comet.exec.aggregate.enabled=false for the affected job.