Remove Tests Subsumed By Stronger Siblings

[chet]

Following #4171, this is the other half of what the test audit turned up: tests that duplicate a sibling already covering the same ground, usually more strictly. The clearest case is libmlx, where several modules grew a tests/ mirror alongside their inline #[cfg(test)] module and the inline one is consistently the stronger of the pair.

What this involves

• Delete only tests whose coverage is provably held by a surviving sibling. Every removal names the test that still covers it -- the PR description is that mapping, and it's what makes a deletion this size reviewable at all.
• Hold the line at high confidence. There's a second tier of maybe-redundant tests I'm deliberately leaving out of this one; those want a per-cluster judgment call rather than a sweep.
• Re-verify every candidate against current main before removing it. The audit predates Fix Tests That Assert The Wrong Thing #4171, and at least one of its targets is already obsolete -- that PR turned the tests in question into real coverage, so deleting them now would lose something.
• The gate is a function-level coverage A/B against main: production coverage has to come out flat. Plain line coverage is worse than useless here, since deleting test lines makes the percentage rise on its own.

Cluster-by-cluster detail lives in the project ledger.

Part of #3914.