Fix FlopsProfiler counter reset between runs

[cklxx]

Summary

• clear global FLOP/MAC collectors when resetting the profiler so each run starts from a clean state
• ensure the collectors are also emptied when ending a profile to avoid leaking stale entries across sessions

Testing

• python - <<'PY'
  import torch
  from torch import nn
  from deepspeed.profiling.flops_profiler import FlopsProfiler

class MyModel(nn.Module):
def init(self):
super().init()
self.lin1 = nn.Linear(4, 4)
self.lin2 = nn.Linear(4, 4)
def forward(self, x):
return self.lin2(torch.relu(self.lin1(x)))

model = MyModel()
inputs = torch.ones(2,4)
profiler = FlopsProfiler(model)
for i in range(3):
profiler.start_profile()
model(inputs)
profiler.stop_profile()
print('iter', i, profiler.get_total_macs(), profiler.get_total_flops())
profiler.end_profile()
PY

---

Codex Task