test: add parser fuzzing and discriminator injectivity checks

[senzenn]

Two properties the generated parsers already have, with nothing asserting them. No
production code changes.

parser_robustness.rs drives every parse entry point with buffers the IDL never described:
every prefix of a valid one including empty, and valid ones with bytes appended. The
contract is return Ok or Err, never panic, which in Rust also rules out an over-read since
a slice past the end aborts. Truncation is the shape that slips through, because a guard
that checks data.len() == N and then slices data[..M] only fails between M and N, which a
happy-path test misses. Random buffers come from a seeded xorshift64* so a CI failure
replays from the seed, and drive returns a tally of outcomes, since "never panicked" is
also true of a corpus that bounced off the first length check every time.

discriminator_injectivity_tests proves no two differently-keyed instructions can both match
one buffer. Equal keys are already handled: the renderer groups them and disambiguates on
account count. Unequal keys get one arm each, tried in order, so an overlap there is decided
by emission order rather than by anything the IDL states. jointly_satisfiable compares two
keys byte-wise over the intersection of their windows, and treats non-overlapping windows as
satisfiable, which is the case that matters. The check runs over the whole corpus rather than
up to a bound.

joint_satisfiability_detects_real_overlap pins the helper on hand-built keys (same window
different bytes, disjoint windows, overlapping and agreeing, overlapping and disagreeing, a
size discriminator against a window past its end), since a helper that always answered no
would make the corpus check pass on any input.
same_key_collision_groups_are_accounted_for pins the four IDLs that do rely on account-count
disambiguation:

colliding_envelope
collision_safe_cpi_event_envelope
macro_arg_envelope_collision
raydium_amm_v4_with_swapv2

so a fifth arrives as a deliberate change.

Neither check depends on the codama bump; both hold on 0.9.1. It sits on #323 to keep the
stack linear.

Tested: 378 passed, 0 failed. clippy -Dwarnings and fmt --check clean.