Use System.IO.Enumeration for faster recursive path traversal

[kthompson]

Why

The old traverser used a manual BFS loop that called GetFiles() / GetDirectories() per directory, allocating DirectoryInfo[] arrays and FileSystemInfo objects for every entry regardless of whether it matched the glob pattern. For sparse patterns like **/*.csproj, this resulted in large allocations (Gen2 GC pressure) and slow traversal.

This PR rewires path traversal to use a single FileSystemEnumerable<FileSystemInfo> with RecurseSubdirectories = true, pushing glob matching directly into the enumeration predicates.

Approach

Two new types do the work:

• PathMatcher -- span-based IsFullMatch and CanRecurse methods that mirror GlobEvaluator's ** backtracking semantics without allocating a string array for path segments.
• RecursiveGlobEnumerable -- wraps a single FileSystemEnumerable with:
  • ShouldRecursePredicate calling PathMatcher.CanRecurse to prune entire subtrees that cannot possibly match the pattern
  • ShouldIncludePredicate calling PathMatcher.IsFullMatch for the full match check
  • Transform (ToFileSystemInfo) only runs on matched entries

Each path is visited exactly once by the OS, so no HashSet dedup is needed. Entries that don't match are never materialized into FileSystemInfo objects.

The one behavioral subtlety: a bare ** pattern also matches the search root itself (zero directories deep). Since FileSystemEnumerable never visits the root, RecursiveGlobEnumerable emits it explicitly when the pattern is all-wildcard.

Removed dead code

With the new traverser in place, the old per-directory machinery is no longer needed:

• Deleted PathTraverserEnumerable.cs (manual BFS)
• TraverseOptions reduced to a plain flags carrier (no GetFiles/GetDirectories, no _dirCache)
• Deleted MockTraverseOptions, MockFileSystemNode, and their unit tests
• Removed System.IO.Abstractions package references from Glob.Tests

Benchmark results (net8.0, Apple M3)

Comparing against origin/develop with identical benchmark definitions ([MemoryDiagnoser] added to baseline for fair comparison):

Benchmark	Baseline	This PR	Time	Alloc
ParseGlob	1,474 ns	1,461 ns	-1%	--
ParseAndCompileGlob	1,495 ns	1,513 ns	+1%	--
MatchForUncompiledGlob	1,759 ns	1,772 ns	+1%	--
MatchForCompiledGlob	199 ns	189 ns	-5%	--
MatchForUncompiledGlobDirectoryWildcard	1,630 ns	1,635 ns	+0%	--
MatchForCompiledGlobDirectoryWildcard	200 ns	190 ns	-5%	--
BenchmarkParseToTree	1,525 ns	1,561 ns	+2%	--
PathTraversal (**/*.cs)	29,635 ns	30,148 ns	+2%	-18%
PathTraversalSparseMatch (**/*.csproj)	1,622,130 ns	714,784 ns	-56%	-70%
PathTraversalDirectories (**/bin)	689,596 ns	535,730 ns	-22%	-46%

Parse/match benchmarks are within noise and allocations are unchanged. Gen2 GC pressure disappears entirely for sparse traversal patterns.

Test coverage

All 179 tests pass (183 minus the 4 MockTraverseOptions unit tests that tested the now-deleted code). Real-filesystem coverage in PathTraverserTests continues to exercise all traversal contracts: dedup, symlink handling, re-enumeration, case-insensitivity, emitFiles/emitDirectories flags, rooted patterns.

Closes #33