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.