[FEA] Raise a warning if LightGBM 64-bit thresholds may cause rounding issues

[JohnZed]

Once treelite supports float64, FIL can detect this and issue a warning.

[JohnZed]

Could possibly do this in the cuML conversion process (before getting into treelite)

[hcho3]

We can add a warning about float64 here:

cuml/python/cuml/fil/fil.pyx

Lines 116 to 144 in a1f9d33

	def from_filename(filename, model_type="xgboost"):
	"""
	Returns a TreeliteModel object loaded from `filename`
	Parameters
	----------
	filename : string
	Path to treelite model file to load
	model_type : string
	Type of model: 'xgboost', or 'lightgbm'
	"""
	filename_bytes = filename.encode("UTF-8")
	cdef ModelHandle handle
	if model_type == "xgboost":
	res = TreeliteLoadXGBoostModel(filename_bytes, &handle)
	if res < 0:
	err = TreeliteGetLastError()
	raise RuntimeError("Failed to load %s (%s)" % (filename, err))
	elif model_type == "lightgbm":
	res = TreeliteLoadLightGBMModel(filename_bytes, &handle)
	if res < 0:
	err = TreeliteGetLastError()
	raise RuntimeError("Failed to load %s (%s)" % (filename, err))
	else:
	raise ValueError("Unknown model type %s" % model_type)
	model = TreeliteModel()
	model.set_handle(handle)
	return model

This covers ForestInference.load() but does not yet cover ForestInference.load_from_treelite_model() or ForestInference.load_using_treelite_handle(). To cover these two, we'd need to bring in the Treelite refactor (dmlc/treelite#196).

[hcho3]

#3316 is needed to reliably detect all Treelite models with float64 values.