Improve efficiency of scheduling and token sampiling

[tohtana]

(NOTE: This PR requires the new APIs introduced in deepspeedai/DeepSpeed#4965)

This PR improves the efficiency scheduling for ragged batching.

• Use a new faster API to query KV cache status (94ebae1)
• Improved efficiency of Top P logits processor by avoiding an inefficient loop (3903024)
• Skip duplicated check of schedulability (904e500)
• Use python int values and lists instead of torch tensors to maintain KV cache status. Using torch tensors has a certain overhead and it becomes significant when they are frequently called in loops (bfdb5db)