Update test cases to prevent platform-specific variability

[ztangent]

Per this comment, there are sometimes situations where Gen's test suite fails due to platform variability (e.g. 32-bit vs 64-bit systems) affecting random number generation even with fixed number generation, and hence approximate equality tests, such as the one below:

Gen.jl/test/inference/particle_filter.jl

Lines 146 to 170 in b959223

	@testset "default proposal" begin
	Random.seed!(1)
	num_particles = 10000
	ess_threshold = 10000 # make sure we exercise resampling
	# initialize the particle filter
	init_observations = choicemap((:x_init, obs_x[1]))
	state = initialize_particle_filter(model, (1,), init_observations, num_particles)
	# do steps
	argdiffs = (UnknownChange(),) # the length may change
	for T=2:length(obs_x)
	maybe_resample!(state, ess_threshold=ess_threshold)
	new_args = (T,)
	observations = choicemap((:chain => (T-1) => :x, obs_x[T]))
	log_incremental_weights, = particle_filter_step!(state, new_args, argdiffs, observations)
	@test length(log_incremental_weights) == num_particles
	end
	# check log marginal likelihood estimate
	expected_log_ml = log(hmm_forward_alg(prior, emission_dists, transition_dists, obs_x))
	actual_log_ml_est = log_ml_estimate(state)
	@test isapprox(expected_log_ml, actual_log_ml_est, atol=0.02)
	end

This platform-specific variability happens despite fixing the random seed to 1. I'm not entirely sure what the most principled way to address this is, but the simple fix would be to relax the tolerance level (e.g. to 0.03 from 0.02) so that platform variability doesn't overly affect the result.