Support pattern matching on Bytes literals - #6228

Overview

This adds support for pattern matching on Bytes literals (0xs...), bringing
them in line with Text, Nat, Int, Float, Char, and Boolean, which
already work as both expressions and patterns.

Old behavior — 0xsabba was accepted in expression position but rejected as a
pattern:

> match 0xs with
    0xsabba -> "abba"
    _       -> "not abba"

I got confused here:
    6 |     0xsabba -> "abba"

I was surprised to find a bytes literal here.
I was expecting one of these instead:
* end of input
* pattern

New behavior — bytes literals can be used as patterns, including
underscore-separated forms (0xs01_ef):

classify : Bytes -> Text
classify = cases
  0xs0102 -> "one-two"
  0xsabcd -> "abcd"
  0xs     -> "empty"
  _       -> "other"

Closes #6128.

Implementation approach and notes

The same shape as the existing Text pattern path, mirrored across every layer:

• AST layers — added Bytes to Unison.Pattern (core),
  Unison.Syntax.Pattern, U.Codebase.Term.Pattern (V2 storage), and
  Unison.Hashing.V2.Pattern. Conversions between layers updated in both directions.
• Parser — added pBytes to pLiteral in TermParser.hs, reusing the
  existing bytesToken lexer.
• Typechecker — Pattern.Bytes requires the scrutinee to be a subtype of
  Type.bytes.
• Pattern-match coverage — added Bytes to PmLit, Vc'Bytes to
  VarConstraints, and threaded it through Solve, NormalizedConstraints, and
  Desugar. Bytes are treated like Text/Char (infinite-domain literals), so
  coverage analysis works without changes to the IntervalSet machinery.
• Pretty-printer — patterns render back as 0xs<hex> for round-trip, matching
  the existing expression-level renderer. bytesRef was added to builtinCase in
  Runtime.Pattern so dispatch goes through splitMatrixBuiltin rather than
  lookupData.
• Runtime — added MatchBytes (Map Bytes e) (Maybe e) to Branched,
  AccumBytes to BranchAccum, TestY/MatchY to MCode, and selectBytesBranch
  in the machine. Dispatch is Map-based (like MatchText), so it's O(log n) rather
  than a chain of equality tests.
• Hashing/serialization — new tags in V2/Pattern.hs (14), Synhash.hs (17),
  MtTag (MBytesT), Sqlite/Serialization (14), MCode/Serialize (TestYT), and
  MurmurHash. Term.intop (which embeds patterns back into terms for some debug
  paths) uses Bytes.fromList [hex...] to mirror how Bytes literal expressions
  desugar.

Interesting/controversial decisions

V4 ANF codec — reject vs. forward-compatible. The V4 codec is a frozen
format predating this feature. I made it explicitly throw on MatchBytes (both
serialize and deserialize) rather than silently fall through. Anyone serializing
with V4 needs to know they've lost a feature. The current codec gets a new tag
MBytesT = 7.

Test coverage

• Added unison-src/transcripts/idempotent/bytes-pattern-match.md covering basic
  dispatch, default fallthrough, the empty 0xs, and underscore-separated literals
  (0xs01_ef). Idempotent — re-running produces zero diff.
• Ran ./scripts/proofs/tests.sh — all unit test suites + 22 CLI integration tests
  pass.
• Ran ./scripts/proofs/transcripts.sh — 385 transcripts pass, no diffs against
  committed .output.md files.
• Coverage feels adequate. Could add explicit pattern-match coverage warnings tests
  (e.g. a non-exhaustive match on Bytes should warn), but that machinery is shared
  with Text/Char so the existing tests for those exercise the same code path.