feature: simple_gla supports sequences of arbitrary length

[pathfinder-pf]

[pathfinder-pf]

1. should we use same kernel for both inference and training? varlen padding or fixed len.
2. gpu inference and training use the same kernel?