Make dtype casts safe

[pavelzw]

Fixes #23

[pavelzw]

This adds safe casts to the old version of the LGBM compression algorithm (as well as the sklearn algorithm). @YYYasin19 needs to do something like this in #15 too.

[github-actions]

(benchmark 4470028469 / attempt 1)
Base results / Our results / Change

Model	Size	Dump Time	Load Time
sklearn rf	77.2 MiB / 11.1 MiB / 6.94 x	0.04 s / 0.11 s / 2.43 x	0.06 s / 0.11 s / 1.76 x
sklearn rf LZMA	17.3 MiB / 5.3 MiB / 3.25 x	38.61 s / 6.69 s / 0.17 x	1.69 s / 0.58 s / 0.34 x
sklearn gb	2.2 MiB / 1.1 MiB / 2.08 x	0.03 s / 0.25 s / 7.47 x	0.04 s / 0.13 s / 3.66 x
sklearn gb LZMA	0.6 MiB / 0.2 MiB / 3.82 x	1.02 s / 0.45 s / 0.44 x	0.09 s / 0.15 s / 1.61 x
LGBM gbdt	5.3 MiB / 1.9 MiB / 2.81 x	0.17 s / 0.49 s / 2.91 x	0.03 s / 0.32 s / 12.11 x
LGBM gbdt LZMA	1.7 MiB / 0.8 MiB / 1.96 x	4.11 s / 1.15 s / 0.28 x	0.16 s / 0.38 s / 2.33 x
LGBM gbdt large	57.6 MiB / 18.9 MiB / 3.05 x	1.72 s / 5.11 s / 2.96 x	0.28 s / 3.55 s / 12.59 x
LGBM gbdt large LZMA	15.3 MiB / 6.7 MiB / 2.29 x	56.13 s / 14.17 s / 0.25 x	1.62 s / 4.00 s / 2.47 x
LGBM rf	1.6 MiB / 0.5 MiB / 3.42 x	0.05 s / 0.09 s / 1.91 x	0.01 s / 0.08 s / 12.93 x
LGBM rf LZMA	0.1 MiB / 0.1 MiB / 1.70 x	0.24 s / 0.13 s / 0.55 x	0.02 s / 0.08 s / 4.22 x

[jonashaag]

Why do we need this special case?