[BUG]Activation recomputation does not support undividable recompute_num_layers

**Describe the bug**
List of out range error when `self.num_layers_per_pipeline_rank % self.config.recompute_num_layers!=0`

**To Reproduce**
Set self.config.recompute_num_layers to a non factor of self.num_layers_per_pipeline_rank

**Expected behavior**
List out of range

**Stack trace/logs**
If applicable, add the stack trace or logs from the time of the error.

**Environment (please complete the following information):**
- Megatron-LM commit ID
- PyTorch version
- CUDA version
- NCCL version

**Proposed fix**
Change [line ](https://github.com/NVIDIA/Megatron-LM/blob/1584dca28bc302102ecc07648f8c286213625c5c/megatron/core/transformer/transformer_block.py#L473) from
`custom(layer_idx, layer_idx + self.config.recompute_num_layers)`
to
`custom(layer_idx, min(self.num_layers_per_pipeline_rank, layer_idx + self.config.recompute_num_layers))

**Additional context**
Add any other context about the problem here.