Document the panics of public functions

[emilk]

Which issue does this PR close?

• Part of Proposal: all panicky functions should have a try_ version #10553

Rationale for this change

This is one of four PRs splitting up the clippy::missing_panics_doc work.
Each one stands on its own and touches its own set of functions, but they are
meant to land in order, since the later ones assume the earlier ones:

1. Return errors instead of panicking in fallible functions #10755 - return errors from fallible functions
2. Remove some unwraps #10759 - remove unreachable panics
3. Document the panics of public functions #10760 - document the panics that genuinely remain
4. Mark unreachable panics with #[expect(clippy::missing_panics_doc)] #10761 - #[expect] the unreachable ones, so the lint can be turned on

What is left after the first two PRs is panics that are real and that the
function cannot report any other way. Those get a # Panics section.

What changes are included in this PR?

• # Panics sections on the public functions that can still panic, including
  the ones that only panic under the force_validate feature
• two of them say why the panic is not an error, so the reader does not have to
  wonder: try_for_each_valid_idx takes the caller's error type and so has no
  error value to build, and the arrow-integration-test JSON readers parse
  fixtures, where malformed input is a broken test rather than bad data
• arrow-csv's Writer::into_inner documents its panic honestly. A successful
  write flushes, but a write that returned an error leaves rows buffered, and
  the flush inside into_inner can then fail. It returns W, so there is
  nowhere to report that
• arrow::util::test_util and arrow::util::bench_util are public, so their
  module docs say the helpers panic on invalid input
• the # Panics sections go before #[inline], #[deprecated] and
  #[proc_macro_derive], rather than after, so the doc block stays in one piece
• closes an unterminated doc code fence in PrimitiveDictionaryBuilder

arrow-integration-testing is deliberately left undocumented: it is
publish = false and holds only integration test binaries. #10761 suppresses the
lint for that crate instead.

Are these changes tested?

Documentation only. cargo doc --workspace --all-features is clean.

Are there any user-facing changes?

Documentation only.