Add various server timeouts, lower --max_batch_size and --inference_max_length defaults

[borzunov]

Summary:

parser.add_argument('--inference_max_length', type=int, default=2048,
                    help='Maximum total sequence length permitted per inference, defaults to 16384 tokens')
parser.add_argument('--max_batch_size', type=int, default=2048,
                    help='The total number of tokens in the same batch will not exceed this value')

parser.add_argument('--alloc_timeout', type=float, default=60,
                    help='If the cache is full, the server will wait for this number of seconds hoping that some memory will be freed '
                         'before rejecting the request')
parser.add_argument('--request_timeout', type=float, required=False, default=3 * 60,
                    help='Timeout for the whole rpc_forward/rpc_backward/rpc_forward_stream/rpc_backward_stream request')
parser.add_argument('--session_timeout', type=float, required=False, default=30 * 60,
                    help='Timeout for the whole inference session')
parser.add_argument('--step_timeout', type=float, required=False, default=5 * 60,
                    help="Timeout for waiting the next step's inputs inside an inference session")

This PR also enables PETALS_8BIT_BACKWARD by default.