[perf] Clear global_vars/matrix_fields after materialize()

[strongoier]

Related issue = #3906

https://github.com/taichi-dev/taichi/runs/4664728269?check_suite_focus=true shows that #3906 causes performance regression. This PR aims at fixing it by clearing global_vars/matrix_fields after materialize() because they should not be handled in the next materialize().

[netlify]

✔️ Deploy Preview for jovial-fermat-aa59dc canceled.

🔨 Explore the source changes: 56debf5

🔍 Inspect the deploy log: https://app.netlify.com/sites/jovial-fermat-aa59dc/deploys/61cd746a6cf96100074fbfb6

[yolo2themoon]

/benchmark

[taichi-gardener]

current commit hash: 56debf5

master commit hash: 0abce04

Examples	CUDA
benchmarks fluctuation threshold	5.0%
kernel improvement	vortex_rings(-27.5858%), mpm128(-26.4935%), comet(-24.8713%), pbf2d(-22.0244%), mpm99(-18.7543%), fem128(-14.9942%), ad_gravity(-13.6921%), mpm88(-12.7889%), odop_solar(-11.7633%), mpm_lagrangian_forces(-11.4748%), fem99(-9.1789%),
kernel regression	inital_value_problem(+7.3329%),
end2end improvement	fem128(-94.2243%), mpm_lagrangian_forces(-93.9893%), fem99(-93.4416%), odop_solar(-88.1788%), pbf2d(-87.6037%), vortex_rings(-87.3154%), mpm128(-86.5948%), comet(-85.8337%), ad_gravity(-84.8365%), mpm99(-83.8713%), mpm88(-83.2098%), mpm3d(-11.4981%), game_of_life(-7.7122%), inital_value_problem(-6.1978%),
end2end regression	None

[yolo2themoon]

LGTM! Thanks!