[dbsp] Fix broken metrics - #6859

A merge marker said what the merge consumed and produced but not what it left
behind? The marker now ends with "spine now holds N loose of M
batches".

The marker moves after the merged batch is added to the spine, so the counts
describe the state the merge leaves rather than the state mid-way through it.

Signed-off-by: Leonid Ryzhyk <ryzhyk@gmail.com>

merge_backpressure_wait_time_seconds read zero on a workload where workers
demonstrably waited for merging, and no backpressure-wait marker appeared in
its profile. There are two ways to wait for the spine to drain and only one of
them was instrumented.

`backpressure_wait` polls in a loop, adds the elapsed time to the spine's
stats, and records a marker. `backpressure_waiter` just hands out a `Notified`
to await, and recorded nothing; the sharded accumulator uses that one
(ShardedAccumulatorLocalWaiter), so its waits reached neither the metric nor
the profile. They were visible only as unexplained time inside the operator.

The waiter now carries what a report needs, its caller reports the wait once
awaited, and both paths funnel into a single `record_backpressure_wait`, so a
wait counts the same however it was taken. The marker reports the batch counts
either side of the wait, loose and total, which says whether waiting achieved
anything.

Splitting the waiter into a future and a report keeps the `Notified` out of the
value the caller holds across the await.

Signed-off-by: Leonid Ryzhyk <ryzhyk@gmail.com>

Every worker shares one `ShardedAccumulator`, built by whichever of them
reaches the constructor first, and that worker's index went to all of the
receive queues. Their compaction then ran on the constructing worker's
background buffer cache and slab allocator, which a spine's worker index
picks, rather than spreading over its peers', and a profile reported all of
the operator's merges under that one worker.

The wrong worker also made concurrent merges of distinct queues look like
repeated merges of a single spine in samply markers: every queue names its
spines after the same operator, so the owner is the only field that tells them
apart.

Construction now takes no worker index at all, so the mistake cannot recur.
A spine carries its owner, so the backpressure-wait marker can name it
alongside the blocked worker that records the wait.

Signed-off-by: Leonid Ryzhyk <ryzhyk@gmail.com>

Its documentation links to the private `BackpressureWait` it is split from,
which rustdoc reports because that link resolves only while private items are
being documented and breaks in the public docs.

The report, and the two `Spine` methods that hand it out and take it back, are
used only by the sharded accumulator inside this crate, so `pub(crate)`
describes them accurately and the link resolves.

Also stop pointing at `RuntimeIdle`'s private `base` field, which rustdoc
flags for the same reason.

`cargo doc -p dbsp` reported eleven warnings that predate this branch, and the
crate's documentation now builds without any.

Four links named items that no longer exist or never did: `StaticScheduler` is
gone, `group_transform` is now `dyn_group_transform`, `size_from_level` is now
`size_to_level`, and `eof` is `BulkRows::at_eof` rather than a method of the
index level whose documentation cited it. Two more spelled out a target that
the link's own label already resolved to, and one wrote `std::mem::transmute`
with a single colon.

The remaining four are prose indexing an array, `weight_times[current_index]`
and `MERGE_COUNTS[level]`, which rustdoc reads as a link. Backticks make them
code, which is what they are.