test(serwalker): unit tests for stream, tags, header and string clusters

[caverav]

What this is

Unit tests for the four layers every cluster sits on: the stream readers, tag
decoding, the snapshot header, and the string cluster. 31 tests, no new
dependencies.

Two commits, kept separate on purpose so the first can be dropped once you
finish the scaffolding.

1. fix(serwalker): make the crate compile again

Heads up: the branch has not built since f50d581, so nine commits back. I
could not run tests against a crate that does not compile, so this commit does
the minimum to get it building. Four things were in the way:

• Stream lost its 'a lifetime parameter when read_bytes_zero_copy landed.
  The struct, the impl block and two signatures stopped resolving. Restored as
  it was.
• walk_snapshot_and_enrich_model used an uninitialised binding, applied ? in
  a function returning ProgramModel rather than Result, and called
  enrich_model_object_pool, which is not exported yet.
• DataSnapshot referenced ProgramRoots, which is not defined anywhere.
• resolve_root_library returned an uninitialised binding.

The unfinished bodies are todo!() now so they typecheck without pretending to
work. Nothing about the intent changed. Drop this commit whenever those are
filled in for real.

2. test(serwalker): cover the stream primitives, tag decoding, header and strings

Each suite encodes its own input the way Dart's WriteStream does, so the tests
assert against the format rather than against our own decoder. A test that just
re-runs the parser would pass with the parser wrong.

suite	tests	what they pin
stream	14	LEB128 round trips for both markers, the Read vs ReadUnsigned distinction, big endian ref ids and the five stage cap, read_byte staying raw, Latin-1 over the full 0..=255 range, bounds checks on every reader
utils	4	ClassIdTag at bit 12, CanonicalBit at 1, ImmutableBit at 7, proven independent of each other, plus unknown cids erroring
snapshot	7	header round trip, the 32 char version and features split, magic and kind rejection, kind numbering with no kModule, counts decoding with the unsigned marker
cluster	6	string cluster round trips for one byte, two byte and mixed clusters, Latin-1 payloads, unpaired surrogates decoding lossily, truncated payload erroring

Three of them pin bugs this PR already hit, so they should stay even if the
surrounding code moves: the_two_markers_are_not_interchangeable,
read_byte_is_raw_not_leb128, ref_ids_are_big_endian_not_leb128.

Proof they would catch a regression

Passing tests on their own do not show much, so I mutation tested them. Fifteen
deliberate regressions, introduced one at a time with the source restored
between each:

mutation	caught by
the two end byte markers swapped	signed_reads_sign_extend
ref ids decoded little endian	ref_ids_are_big_endian_not_leb128
the +128 ref id correction dropped	read_ref_id_round_trips
five stage ref id cap removed	ref_id_is_bounded_to_five_stages
Latin-1 reverted to from_utf8	latin1_accepts_the_high_half
bounds check removed from take	truncated_input_errors_instead_of_panicking
read_byte turned into a LEB read	compile break
each of the three tag shifts moved by one	the_flag_bits_do_not_overlap
version split shortened to 31	splits_version_from_features_at_thirty_two_chars
num_objects read with the signed marker	header_counts_use_the_unsigned_marker
magic check disabled	rejects_a_bad_magic
a seventh snapshot kind accepted	snapshot_kind_numbering_matches_dart
string length shift dropped	one_byte_strings_carry_latin1_payloads
from_utf16_lossy made strict	unpaired_surrogates_do_not_panic

Fifteen out of fifteen. One test was weak on the first pass:
ref_id_is_bounded_to_five_stages originally used six truncated bytes, so it
was really testing truncation and survived the cap being removed. It now uses
eight continuation bytes followed by a valid terminator, so only the cap can
reject it.

One source change outside the tests

DECLARE_FIXED_LENGTH_CLUSTER and DECLARE_VARIABLE_LENGTH_CLUSTER now derive
Default on the generated struct. Without it a cluster can only be built
through decide_cluster, which still returns Err for nearly everything, so
_StringCluster was untestable. Two lines, no behaviour change.

Verifying locally

cargo test -p flutterdec-serwalker
cargo fmt --all --check

Current state on this branch: fmt clean, builds, 31 passed 0 failed.

Not included

Clippy still reports 18 findings, none of them in the test code. They are the
unfinished scaffolding: unused imports in info_producer, std::env::var,
std::default, anyhow::Error, and five u8 as u8 casts left over from the
read_byte conversion. Left alone so this PR stays about tests.