Fix silent string truncation in PyGrain shared memory batching

[copybara-service]

Fix silent string truncation in PyGrain shared memory batching

The prompt truncation is caused by a silent data-loss bug in PyGrain's shared memory batch stacking implementation which is active during evaluation when num_workers > 0 (multi-process prefetching).

To improve multi-process data loading performance, Kauldron automatically enables output_to_shared_memory=True inside PyGrain's batching transformations.

When batching elements inside shm_stacking_function, the destination SharedMemoryArray's shape and dtype are inferred exclusively from the first element in the batch:

first_arg = np.asanyarray(args[0])
shape, dtype = (len(args),) + first_arg.shape, first_arg.dtype

In a batch containing variable-length string elements, the SharedMemoryArray is allocated with a fixed-width unicode dtype bounded by that first string.

When np.stack(args, out=...) copies subsequent longer strings into the allocated fixed-width shared memory array, NumPy silently truncates the strings to the array's fixed width!