Add relative tolerance to the rtp limit comparison

[soulrap]

Problem

verify_mode_volatility compares each statistic against its limit with a strict
val > limit. The optimizer targets the rtp limit exactly, so its result can
land a few ulps on either side of it. When it lands above, the run is reported as
failing a limit it actually hit.

Two of the sample games reproduce this on a clean full run
(run_sims + run_optimization + run_analysis + run_format_checks, 1e4 sims per mode):

VIOLATED: rtp  VALUE:0.967 --- LIMIT: 0.967

The printed value is round(val, 4), which hides the cause. The exact values,
recomputed from the published lookup tables:

game	mode	rtp	result
0_0_expwilds	superspin	0.96700000000123722188	flagged (+1.2e-12)
0_0_scatter	bonus	0.96700000000008785861	flagged (+8.8e-14)
0_0_expwilds	base	0.96699999947532203403	passed
0_0_scatter	base	0.96699999991445092640	passed

Which side of the limit a mode lands on is incidental to the optimizer's weight
search, so the failure is not reproducible in any meaningful sense.

Change

Compare against the limit with a relative tolerance, applied to rtp only:

limit_rel_tol = {"rtp": 1e-9}

for key, limit in mode_limits.items():
    val = getattr(MathStats, key, None)
    if val is None or val <= limit:
        continue
    if isclose(val, limit, rel_tol=limit_rel_tol.get(key, 0.0)):
        continue
    violated_attributes[key] = val

Three decisions behind it:

1. rtp only. Limits absent from limit_rel_tol get rel_tol=0.0, and since
   val > limit already excludes equality, that is exactly the previous strict
   behaviour. Only rtp is a value the optimizer deliberately drives onto the
   boundary, so only rtp has a reason to be loosened.
2. Relative, not absolute. 1e-9 relative is ~9.7e-10 absolute at 0.967 and
   rescales on its own if the limit is ever retargeted (0.97, 0.95, ...).
3. 1e-9 in size. The largest residue observed was 1.2e-12, leaving ~800x
   headroom, while the smallest meaningful rtp regression (0.9671) sits five
   orders of magnitude away. The tolerance cannot mask a real deviation.

Verification

Six cases through verify_mode_volatility:

scenario	rtp	expected	result
0_0_expwilds/superspin actual	0.96700000000123720	pass	pass
0_0_scatter/bonus actual	0.96700000000008790	pass	pass
exactly at limit	0.967	pass	pass
below limit	0.96699999947532200	pass	pass
real violation	0.9671	VIOLATED	VIOLATED
unoptimized run	17.7082	VIOLATED	VIOLATED

Other metrics keep the strict comparison: cvar = 800.0000000001 against a limit
of 800 is still reported as violated.

execute_all_tests was then re-run against the existing publish files for both
affected games. All five modes report SHA-256 OK, payout hash OK with no
VIOLATED line.

All six sample games were run end to end with optimization enabled (13 modes
total, every one landing on rtp 0.967) to confirm nothing else regressed.