[BUG] deepspeed_checkpoint.get_transformer_state has more PP_degree than real PP_degree

Describe the bug

when use deepspeed_checkpoint get transfromer_state, self.transformer_file_map has more pp_degree than real

To Reproduce

original code

consider a transformer model with 70 transformer-layer, PP = 12

original code will set layers_per_pp as 70 // 12 = 5
now, transformer_file_map will have pp_degree as 70 // 5 = 14

Possible Fix

def _build_transformer_file_map(self):
        transformer_layer_keys = self.layer_keys[1:-1]
        file_map = {}
        # XXX: this is not guaranteed
        import math
        layers_per_pp = math.ceil(len(transformer_layer_keys) / self.pp_degree)
        if layers_per_pp == 0:
            layers_per_pp = 1
        ...

Start in deepspeed/checkpoint/deepspeed_checkpoint.py at _build_transformer_file_map and inspect how transformer_layer_keys and pp_degree determine the map. Reproduce the 70-layer, PP=12 case and verify that the resulting transformer_file_map does not contain more pipeline stages than requested.