Fix big chunk leak in allocator (#26017)

Fix proposed by GPT 5.6 Sol.

close #26016
Potentially close #22510

No concrete proof for the second one, though the described behavior
matches and the step count explains why it's so difficult to find a
repro.

(cherry picked from commit 3bb46d3217)

tests/threads/tsharedfreelistbigchunks.nim

@@ -0,0 +1,40 @@
+discard """
+  matrix: "--mm:arc; --mm:orc"
+"""
+
+import std/[atomics, typedthreads]
+
+const numChunks = 23 # More than the allocator's bounded drain can process.
+
+var
+  pointers: array[numChunks, pointer]
+  allocated: Atomic[bool]
+  continueAllocating: Atomic[bool]
+
+proc allocPointers() {.thread.} =
+  for i in 0..<pointers.len:
+    pointers[i] = allocShared(8192)
+  allocated.store(true, moRelease)
+
+  while not continueAllocating.load(moAcquire):
+    discard
+
+  # The first allocation drains MaxSteps + 1 chunks. The second allocation
+  # must still be able to find and drain the remainder.
+  for _ in 0..1:
+    let p = allocShared(8192)
+    deallocShared(p)
+
+  doAssert getOccupiedMem() == 0
+
+var thread: Thread[void]
+createThread(thread, allocPointers)
+
+while not allocated.load(moAcquire):
+  discard
+
+for p in pointers:
+  deallocShared(p)
+continueAllocating.store(true, moRelease)
+
+joinThread(thread)