Run all three paragraph directions in the BidiTest conformance suite

[youdie006]

This does not fix a library bug — unicode-bidi is fully conformant. It fixes
the conformance runner, which has been asserting a third of BidiTest.txt and
reporting success. I want that stated first so the PR is judged as what it is.

The predicate

gen_base_levels_for_base_tests (tests/conformance_tests.rs:204):

.filter(|bit| bitset & (1u8 << bit) == 1)

can only ever be true for bit == 0. For bit == 1 the expression is
bitset & 2, which is 0 or 2; for bit == 2 it is bitset & 4, which is 0
or 4. Neither is ever 1. So VALUES[1] (LTR) and VALUES[2] (RTL) are
unreachable.

It contradicts its own surroundings: the doc comment on line 200 names all three
values ("auto-LTR, LTR, RTL"), the assert on line 202 requires all three bits to be
meaningful, and tests/data/BidiTest.txt:71 documents the field as
1 = auto-LTR, 2 = LTR, 4 = RTL.

What that costs

Counted two independent ways — a script over the data file, and an AtomicUsize
instrumenting the loop, which agreed exactly:

data lines in BidiTest.txt	490,846
conformance cases (sum of popcount(bitset))	770,241
cases asserted today	256,747 (exactly 1/3.00)
skipped	513,494 (66.7%)
data lines producing zero assertions	234,099 (47.7%)

The bitset distribution in the file is {2: 51303, 3: 182796, 4: 182796, 5: 51303, 7: 22648} — no line has bitset 1, so every line with bitset 2 or 4 currently
produces an empty vector, the for body never runs, and passed_num never
increments, so the panic at :179 cannot fire.

The runtime is corroborating evidence: cargo test --release --test conformance_tests goes from 0.75s to 1.87s on an unchanged library.

The change

== 1 → != 0, plus a unit test for the helper.

The test rows pin the boundary from both sides: bitset 1 pins bits 1 and 2
excluded when unset, bitsets 2 and 4 pin them included, bitset 5 pins a
non-contiguous mask so an off-by-one cannot survive, and bitset 7 pins ordering.
Reverting the predicate fails it with left: [], right: [Some(Level(0))].

Cost to CI, up front

The test suite gets about 2.5–3x slower — that is the point, but you should
hear it from me rather than notice it. Prebuilt binaries, alternating order, 5
reps, plus a byte-identical second build of master as a noise control:

	min	median
master	3.85s	4.34s
master, rebuilt (noise control)	4.28s	4.63s
this PR	10.43s	10.60s

Noise floor +11.2% min; this PR +170.9% min. It lands on the 770241/256747 = 3.00
case-count ratio, which is what you would expect if the work simply was not being
done before. Roughly +7s per test row.

Verified rows

cargo test, --features serde, --no-default-features --features hardcoded-data,
and MSRV 1.47.0 all pass with the fix (4 passed each). cargo fmt --check is
clean. I also checked --no-default-features explicitly rather than assuming:
[[test]] required-features = ["hardcoded-data"] means the conformance test is not
built on that row at all.

Two related gaps I found and did not include

Both are real, both currently pass, and I left them out so this diff stays one
thing. Happy to add either here or separately:

1. exp_base_level is parsed and never asserted. BidiCharacterTest.txt field
   2 is read at :237 and only stored into the Fail struct at :273, while
   actual_base_level is hardcoded None at :278. I added the missing assertion
   against para.level: 0 mismatches across all 91,709 lines.
2. test_character_conformance never exercises the UTF-16 API, while
   test_basic_conformance does at :138-145 (added in f04d399, never mirrored
   to the sibling). I cross-checked BidiInfoU16 levels, paragraph level,
   reorder_visual and has_rtl across all 91,709 lines: 0 divergences.
   Corrupting a UTF-16 unit as a control fired on 47,082 / 20,596 / 9,863 / 17 lines
   respectively, so that zero is real. Worth knowing it buys less than it sounds
   like: BidiCharacterTest.txt contains no astral code points, so no
   surrogate-pair path is covered either way.

Also worth flagging separately

src/char_data/tables.rs:8 says UNICODE_VERSION = (17, 0, 0), but both vendored
conformance files are Unicode 15.0.0 (BidiTest-15.0.0.txt,
BidiCharacterTest-15.0.0.txt). I did not touch that here — refreshing the fixtures
is a separate change, and it is the place where a genuine library divergence would
most plausibly turn up.

Disclosure

I used an AI assistant to help find and prepare this change. I reviewed and tested
it myself, and the counts and timings above are measurements I ran.