[frankmcsherry]

inspect no longer observes frontiers, so it no longer keeps an enclosing region awake, and a new inspect_stamped does the same for stamped containers.

inspect was built on inspect_core, which is scheduled whenever its input frontier changes.
A region takes the strongest frontier interest among its children, so one inspect inside a region means the region is scheduled every time time moves, and then does progress work for all of its children.
This PR adds inspect_stamped(|stamp, container| ..), built on unary, which declares no frontier interest, so it is scheduled only when containers arrive.
inspect is built on inspect_stamped.
inspect_core still observes frontiers, and its documentation now says what that costs; it was previously the only way to see stamps.

Data.
N dataflows of input -> region { map^L -> inspect } -> probe, one record into one input per round, all inputs advanced, one worker, release build.
On master inspect is inspect_core underneath, so the left column is master's behavior.

N x L	inspect_core (master's inspect)	inspect (this PR)
100 x 100	2,226 µs/round	184 µs/round
1,000 x 100	24,772 µs/round	1,444 µs/round
100 x 1,000	24,832 µs/round	942 µs/round

This came up in a spreadsheet engine built on differential dataflow (see #833), which observes each formula's output with inspect.
With its 2,000 formulas grouped into regions, an edit that reaches one formula took 1.75 ms per edit with frontier-observing inspects, and 0.07 ms with this change.

The new test counts how often a region holding only an inspect operator is scheduled as an input advances 100 rounds without data: 104 times with inspect_core, and 5 times with either inspect_stamped or inspect.

🤖 Generated with Claude Code

[antiguru]

Looks good, thank you!