Use multi-dimensional arrays to represent states

[giordano]

This implements one of the possible improvements suggested in #53. However, I don't see any actual improvement in memory allocations compared to what was reported there:

julia> tdac(TDAC.tdac_params(;nprt = 1, n_time_step = 1, n_integration_step = 251, time_step = 251.0, nobs = 4, enable_timers = true));
 ────────────────────────────────────────────────────────────────────────────────
                                         Time                   Allocations      
                                 ──────────────────────   ───────────────────────
        Tot / % measured:             465ms / 100%             667MiB / 100%     

 Section                 ncalls     time   %tot     avg     alloc   %tot      avg
 ────────────────────────────────────────────────────────────────────────────────
 Particle State Update        1    183ms  39.4%   183ms    306MiB  46.0%   306MiB
 True State Update            1    167ms  35.9%   167ms    306MiB  46.0%   306MiB
 Initialization               1   92.7ms  20.0%  92.7ms   43.6MiB  6.54%  43.6MiB
 Process Noise                1   20.4ms  4.40%  20.4ms   8.55MiB  1.28%  8.55MiB
 Particle Variance            1    668μs  0.14%   668μs   1.83MiB  0.27%  1.83MiB
 Resample                     1    265μs  0.06%   265μs     96.0B  0.00%    96.0B
 Particle Mean                1    238μs  0.05%   238μs     0.00B  0.00%    0.00B
 State Copy                   1   77.6μs  0.02%  77.6μs     32.0B  0.00%    32.0B
 Weights                      1   14.0μs  0.00%  14.0μs   1.03KiB  0.00%  1.03KiB
 Observations                 2   11.3μs  0.00%  5.66μs      224B  0.00%     112B
 Observation Noise            1    827ns  0.00%   827ns     48.0B  0.00%    48.0B
 ────────────────────────────────────────────────────────────────────────────────

Nevertheless, I think that having the states represented with multi-dimensional arrays simplifies the code in many places as it allows us to avoid doing complex index arithmetic, we don't need dim_grid and dim_state at all, and we don't need to flatten arrays in a few places.

[tkoskela]

The d_create function in HDF5.jl was so weird that I have to check what this actually does. It looks like it may be creating a dataset with the size of all the variables, and we only want one.