get_resource_semaphore: add interval_seconds beside sample_interval_seconds (#3653 item 17)

[erikdarlingdata]

Part of #3653 (item 17 remainder).

What changed

Item 17's window_truncated rename shipped earlier (#3793). One crossed spelling remained: the fact
"the interval the deltas cover, null when unknowable" is published as interval_seconds by
get_latch_stats / get_spinlock_stats (both products), but as sample_interval_seconds (+
interval_known) by get_resource_semaphore. Per the house rule (add a wire field beside, never rename),
this PR adds interval_seconds next to sample_interval_seconds on get_resource_semaphore, on both
products:

• Darling: Darling/PerformanceMonitor.Darling.Service/Mcp/DarlingMcpMemoryGrantTools.cs
• Lite: Lite/Mcp/McpMemoryTools.cs

interval_seconds carries the exact same value as sample_interval_seconds
(r.SampleIntervalSeconds is > 0 ? r.SampleIntervalSeconds : null), including null for a restart-marker
row (stored 0) and a pre-column row (stored NULL). sample_interval_seconds and interval_known are
untouched. The two sites that publish the RAW stored column (0 when unknown) — DarlingMcpTrendTools.cs:188
and Lite/Mcp/McpPerfmonTools.cs:125 — were not touched; they're a different fact (the raw column, not the
null-when-unknowable derived one).

Description tail (after <<GUIDE>>, never the head): one sentence added to both products' identical
tails, naming the parity: "interval_seconds is the interval these deltas cover, null when it can't be
known — the same name the latch and spinlock tools use; sample_interval_seconds carries the same value and
is kept for existing readers." The heads (before <<GUIDE>>) are untouched and stay byte-identical between
products, so the tools/list budget pins (McpToolsListBudget/DarlingMcpMemoryGrantTools.txt and
McpToolsListBudget/McpMemoryTools.txt, get_resource_semaphore line) are unaffected — served length is
617 on both, unchanged. No #nnnn in the added text. These files sit outside the #3898 description fence,
and no #3898 PR is open (checked before editing).

Tests: extended the existing live-Postgres round-trip
(Darling.Tests.DarlingMcpMemoryGrantToolsLivePostgresTests, which plants a restart marker, a measured
(120s) row, and a pre-column NULL row across three pools) with interval_seconds assertions matching
sample_interval_seconds for all three, including null for the two unknowable cases. Added the mirrored
assertion on Lite's unit test (Lite.Tests.McpLatestSnapshotStampTests.GetResourceSemaphore_ReadsTheWindowBesideTheStampedSnapshot),
which seeds pre-column rows, asserting both fields read null there. No twin-parity census covers this
tool's payload keys directly, so no such file needed updating.

What I ran

• dotnet build Darling/Darling.Tests/Darling.Tests.csproj -p:EnableWindowsTargeting=true — clean.
• dotnet build Lite.Tests/Lite.Tests.csproj — clean (Lite.Tests does not execute on macOS; build-only per lane rules).
• Targeted classes on Darling.Tests (Windows-desktop stripped from runtimeconfig for the run, restored after):
  DarlingMcpMemoryGrantToolsSurfaceAndSqlTests (13 passed), McpToolsListBudgetTests (5 passed, run 3x for
  stability), DarlingMcpMemoryGrantToolsLivePostgresTests against a throwaway
  timescale/timescaledb:2.28.1-pg18 container (port 55893, removed after) — 1 passed.
• Full Darling.Tests suite, once: 13307 total, 214 failed, 714 skipped — the same 214 Windows-only
  failures reported as the dev baseline on this Mac (DarlingInstallLocationTests, TargetProviderTests,
  ViewerDrillDownTests, etc. — none in DarlingMcpMemoryGrantTools* or McpToolsListBudgetTests).

What I did NOT verify

• Lite.Tests execution (WPF; CI-only per lane rules).
• No live-Postgres run on the Lite side (Lite's twin test here is a DuckDB unit test, not gated on
  DARLING_TEST_PG).

CHANGELOG entry

• get_resource_semaphore (Darling and Lite) now also publishes interval_seconds, the same
  null-when-unknowable interval fact get_latch_stats and get_spinlock_stats already use under that name,
  beside the existing sample_interval_seconds (kept for existing readers).