Precision loss for various `taichi` kernels? (f64 -> f32)

[EricCousineau-TRI]

Super minor, but one such example, off of ac1a2f7:

$ python3 -m plb.algorithms.solve --algo action --env_name Rollingpin-v1 --path output
...
[Taichi] version 0.7.26, llvm 10.0.0, commit e37bdb5e, linux, python 3.6.9
[Taichi] Starting on arch=cuda
...
[W 08/01/21 19:56:59.773 25212] [type_check.cpp:visit@70] [$10142] Atomic add (f32 to f64) may lose precision, at
[W 08/01/21 19:56:59.773 25212] [type_check.cpp:visit@71]
...
[W 08/01/21 19:57:02.387 25212] [type_check.cpp:visit@100] [$17705] Local store may lose precision (target = f32, value = f64) at
[W 08/01/21 19:57:02.387 25212] [type_check.cpp:visit@101]
...
  File ".../PlasticineLab/plb/engine/mpm_simulator.py", line 198, in grid_op
    v_out = self.primitives[i].collide(f, I * self.dx, v_out, self.dt)
  File ".../PlasticineLab/venv/lib/python3.6/site-packages/taichi/lang/kernel_impl.py", line 50, in decorated
    return fun.__call__(*args)
  File ".../PlasticineLab/venv/lib/python3.6/site-packages/taichi/lang/kernel_impl.py", line 122, in __call__
    ret = self.compiled(*args)
  File ".../PlasticineLab/plb/engine/primitive/primive_base.py", line 93, in collide
    dist = self.sdf(f, grid_pos)
  File ".../PlasticineLab/venv/lib/python3.6/site-packages/taichi/lang/kernel_impl.py", line 50, in decorated
    return fun.__call__(*args)
  File ".../PlasticineLab/venv/lib/python3.6/site-packages/taichi/lang/kernel_impl.py", line 122, in __call__
    ret = self.compiled(*args)
  File ".../PlasticineLab/plb/engine/primitive/primive_base.py", line 59, in sdf
    grid_pos = inv_trans(grid_pos, self.position[f], self.rotation[f])
  File ".../PlasticineLab/venv/lib/python3.6/site-packages/taichi/lang/util.py", line 163, in wrapped
    return func(*args, **kwargs)
  File ".../PlasticineLab/venv/lib/python3.6/site-packages/taichi/lang/impl.py", line 202, in maybe_transform_ti_func_call_to_stmt
    return ti_func(*args, **kwargs)
  File ".../PlasticineLab/venv/lib/python3.6/site-packages/taichi/lang/common_ops.py", line 227, in assign
    return ti.assign(self, other)
  File ".../PlasticineLab/venv/lib/python3.6/site-packages/taichi/lang/ops.py", line 130, in wrapped
    return a.element_wise_writeback_binary(imp_foo, b)
  File ".../PlasticineLab/venv/lib/python3.6/site-packages/taichi/lang/matrix.py", line 179, in element_wise_writeback_binary
    ret.entries[i] = foo(self.entries[i], other.entries[i])

Once again, great work!!!

[zongong]

My Taichi version 0.7.8 runs.