Accumulate n_tokens_seen from fwd/bwd step to calculate MFU/Throughput

[ighoshsubho]

• Fixed accum ntokens_since_last_log of n_tokens_seen from forward_backward_step

• The MetricsProcessor calculates it like so

time_delta = time.perf_counter() - self.time_last_log

# tokens per second per device, abbreviated as tps
tps = self.ntokens_since_last_log / (
    time_delta * self.parallel_dims.non_data_parallel_size
)
# model FLOPS utilization
# For its definition and calculation, please refer to the PaLM paper:
# https://arxiv.org/abs/2204.02311
mfu = 100 * self.num_flops_per_token * tps / self.gpu_peak_flops
tflops = self.num_flops_per_token * tps / 1e12

[dmahan93]

LGTM! Thank you!