fix: handle all test_range types in SizePartitioner.get_cost

[octo-patch]

Fixes #2430

Problem

SizePartitioner.get_cost() computed actual_size using an eval trick:

actual_size = eval('len(range(self.dataset_size[dataset_abbr]))'
                   f'{test_range})')

This constructs expressions like len(range(N)[:100]) and relies on test_range being a string slice. However, test_range can also be None, an int, or a float (all documented in DatasetReader), which causes a SyntaxError:

• None → len(range(N))None) — syntax error
• int 100 → len(range(N))100) — syntax error
• float 0.5 → len(range(N))0.5) — syntax error

Solution

Add _get_actual_size(test_range, total_size) that mirrors the branching logic already present in load_partial_dataset() (in icl_dataset_reader.py) without loading the dataset into memory:

test_range type	Behaviour
None / empty string	full dataset size
int / float outside (0, total_size)	full dataset size
float in (0, 1)	int(test_range * total_size)
int in (0, total_size)	int(test_range)
str (e.g. "[:100]")	len(index_list{test_range})

Both eval call-sites in get_cost() are replaced with calls to this helper.

Testing

Manually verified with representative inputs:

p._get_actual_size(None, 1000)      # 1000
p._get_actual_size('', 1000)        # 1000
p._get_actual_size(100, 1000)       # 100
p._get_actual_size(0.5, 1000)       # 500
p._get_actual_size('[:100]', 1000)  # 100
p._get_actual_size('[100:200]', 1000) # 100