Express requirements as carets, and take ps-reclaim 0.1.1

[pathscale]

Every = requirement in Cargo.toml becomes a caret, ps-reclaim moves to the
release that fixes two use-after-free windows, and the version goes to
1.0.0-beta.17.

Requirements

dependency	before	after
data_bucket	"=0.5.5"	"0.5.5"
indexset (WorkTablesIndex)	"=0.0.9"	"0.0.9"
vanilla_indexset (indexset)	"=0.15.0"	"0.15.0"
worktable_codegen	"=1.0.0-beta.15"	"1.0.0-beta.15"
ps-reclaim	"0.1.0"	"0.1.1"

Two of these are worth being precise about.

ps-reclaim was never locked. A bare "0.1.0" is already ^0.1.0, so
>=0.1.0, <0.2.0, and 0.1.1 satisfies it: beta.16 picks up the fix on any
fresh resolve without this change. The bump raises the floor above the version
with the Send guard, so a resolve cannot land on it at all. That is the
reason to change the line, not the caret.

Caret on a 0.0.x version buys nothing. ^0.0.9 means >=0.0.9, <0.0.10,
which is the same set =0.0.9 describes. WorkTablesIndex needs a 0.1.0
release before the requirement can widen. The = is gone either way, so the
file no longer implies a constraint it is not expressing.

worktable_codegen is the one that carries risk. The exact pin was
enforcing lockstep between the macro and the runtime it generates calls into,
and ^1.0.0-beta.15 now admits beta.16, beta.17 and 1.0.0. A mismatched pair
fails at expansion rather than at resolve, so it will be a compile error in a
consumer rather than a resolver conflict here. Changed as asked; noted once.

ps-reclaim 0.1.1

Published, and pathscale/ps-reclaim now has CI (build, test, doctests, fmt,
clippy, Miri under strict provenance, and publish from master) where it
previously had no workflows at all. 0.1.1 fixes:

• Guard was Send while its own documentation said it was not, so dropping a
  sent guard stored NO_DOMAIN into the originating thread's pin slot and
  decremented the wrong thread's DEPTH. It is now !Send by construction,
  held by a compile_fail doctest.
• Guard was three words against crossbeam-epoch's one. partition_ref
  returns one per lookup, so it paid the size on every call. It is now one word.

docs/TODO.md

The two items that were listed as blocking beta.16 and did not stop it are
moved to a "Closed, and how" section: the ps-reclaim publish, and CI having
never run on that branch. What remains blocking is re-measuring the partition
regression on a quiet machine, now against 0.1.1 rather than against the guard
size that 0.1.1 removes, and deciding whether to yank ps-reclaim 0.1.0 and
beta.16.

The file also now says explicitly not to use the second set of partition_ref
numbers that circulated during the beta.16 release. They were taken before
673869c showed the benchmark arm labelled pinned_get was calling
partition_ref, they were taken under load, and they disagree with the
interleaved run by roughly a factor of two on beta.15. Neither set has been
reproduced on a quiet box.

Verification

cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings and cargo test --workspace --all-targets all clean locally against
ps-reclaim 0.1.1 resolved from the registry. Cargo.lock is gitignored here,
so CI resolves fresh.

Merging cuts 1.0.0-beta.17, since the publish job releases whatever version
master carries. The perf item above is open at merge time.

[pathscale]

Superseded by #87, which carries this commit rebased onto master. One change on top: the two intra-workspace pins (worktable_codegen, and now worktable_dsl) go back to =, for the reason this PR's own description gave. External requirements stay carets.