fix: expand aliased slice views with differing lengths in TagExpand

[linkdata]

Fixes #203.

Problem

TagExpand's cycle detector (sameActiveNode) identified an active []any
node by its first-element pointer alone. For a slice, reflect.Value.Pointer()
is the address of the first element, so two distinct slice headers viewing the
same backing array at the same start compared equal even with different lengths.

Expanding a short view whose element references the longer backing slice then
mistook the longer slice for a cycle back to the short view, skipped its
expansion, and silently dropped every later tag:

all := make([]any, 3)
outer := all[:2]
all[0] = tag.Tag("first")
all[1] = all
all[2] = tag.Tag("last")
tag.TagExpand(nil, outer) // returned {"first"}; "last" dropped with no error

Downstream Dirty/broadcast/lookup on the dropped tag then silently miss those
Elements.

Fix

Give reflect.Slice its own case in sameActiveNode that also compares length:

• Views of differing lengths iterate different elements, so they are distinct
  traversal nodes.
• A genuine self-referential slice re-enters with an identical pointer and
  length and is still detected as a cycle.

Capacity is intentionally not compared — two views with the same start and
length iterate identical elements, so they are correctly the same node. The
change only makes detection more precise (removes a false positive); even a
hypothetical missed cycle stays bounded by maxTagDepth/maxTagCount.

Tests

• TestTagExpand_AliasedSliceViews — the issue's exact reproduction.
• TestSameActiveNode_AliasedSliceViews — direct helper coverage of the
  differing-length (distinct) and self (same-node) cases.

Both fail without the fix and pass with it. Existing cycle tests
(TestTagExpand_SelfReferentialSliceStopsRecursing, the TagGetter self/mutual
cycle tests) stay green.

Verified with gofmt -l, go vet ./lib/tag/, and go test -race ./...
(all packages green, no races).

Known follow-up (out of scope)

FindTagGetter in lib/tag/errnotusableastag.go has an analogous
seenPtr{t, ptr} slice-aliasing collision, but it only powers a best-effort
error-message hint — over-detection there degrades a hint string, never drops
real tags. Left untouched to keep this change focused on the reported bug.