…thread

Covers the regression from issue ClickHouse#102183: on builds without
JEMALLOC_BACKGROUND_THREAD (e.g. macOS), je_mallctl for "background_thread"
returns ENOENT. The new gtest exercises three guarantees from the fix:

  - tryGetValue reports failure for an unknown mallctl and does not
    write through the output pointer.
  - getValue returns a zero-initialized value when the mallctl is missing.
  - verifySetup does not abort or warn when the only potential mismatch
    comes from optional background-thread mallctls that are absent on
    the running jemalloc build.

Per review (ClickHouse#104330): collapse the je_mallctl boilerplate by implementing
getValue on top of tryGetValue, and chassert that the read succeeded.
This makes the contracts explicit:

  - tryGetValue is the fallible API for optional mallctls.
  - getValue means "this mallctl must exist; failing is a programming
    error" and trips chassert in debug if violated.

Drop the now-incompatible "GetValueReturnsZeroOnUnknownMallctl" gtest
since calling getValue on a missing mallctl is no longer defined.
The remaining tests still cover the regression from ClickHouse#102183.

Previously implemented Jemalloc::getValue on top of tryGetValue and
chasserted that the read succeeded. CI for PR ClickHouse#104330 showed that this
contract aborts the server at startup: AsynchronousMetrics::update reads
jemalloc stats via getValue, and on builds where any individual mallctl
is unavailable (e.g. background-thread mallctls on macOS, or any future
optional name) the chassert fires inside saveJemallocMetricImpl and
debug builds abort with signal 6.

Switch back to a non-fatal contract that better matches the call sites:

  - getValue returns a zero-initialized value when je_mallctl fails.
    AsynchronousMetrics will report 0 for an unreadable metric instead
    of aborting the entire server.
  - tryGetValue stays the explicit API for callers that must distinguish
    "missing" from "present" (verifySetup already uses it).
  - MibCache::getValue zero-initializes its local for the same reason,
    so a failing je_mallctlbymib no longer returns uninitialized memory.

Restore the gtest pinning the zero-default contract for getValue, which
was dropped in 2da2008 when that contract was briefly undefined.

Closes ClickHouse#102183.

saveJemallocMetricImpl now uses Jemalloc::tryGetValue and only inserts
into AsynchronousMetricValues when je_mallctl succeeds. Optional or
build-dependent stats (e.g. background_thread.* on macOS) no longer
appear as misleading zeros.

jemalloc.epoch is only published when epoch can be read without forcing
an update (same tryGetValue rule).

MergeTree arena derived bytes metrics are emitted only when both
pactive and pdirty reads succeed.

chassert when je_mallctl fails, so absent mallctl names are only used where
call sites explicitly tolerate failure.

Use tryGetValue for best-effort reads that must not abort debug builds:
MemoryTracker (prof.active, opt.prof_prefix when flushing profiles),
MemoryWorker (arenas.narenas before per-arena dirty_decay_ms), and
KeeperJemallocStatusHandler (opt.prof, prof.active, prof.lg_sample for JSON).
Remove the try/catch around the old getValue path there because getValue
does not throw.

AsynchronousMetrics jemalloc stats already used tryGetValue.

Drop the gtest that expected getValue to return zero for a bogus mallctl;
that conflicts with the strict getValue contract.

…o fix-jemalloc-verify-setup-darwin

…o fix-jemalloc-verify-setup-darwin

…o fix-jemalloc-verify-setup-darwin