[VM] Move param bind to OptimizeModule

[masahi]

In VM, BindParamsByName is run inside VMCompiler::Lower(...) before OptimizeModule gets called. So when we try to dump the optimized module like this:

compiler = relay.vm.VMCompiler()
opt_mod, _ = compiler.optimize(mod, target, params=params)
print(opt_mod)

param binding doesn't happen, so the "optimized graph" has parameters as arguments. This means constant folding involving parameters cannot happen, so the optimized graph looks very different (e.g. there are many layout_transform) from the output of relay.optimize, which does param binding when relay.optimize is called.

tvm/src/relay/backend/build_module.cc

Lines 255 to 263 in 384714b

	IRModule Optimize(IRModule relay_module, const TargetsMap& targets,
	const std::unordered_map<std::string, runtime::NDArray>& params) {
	ICHECK(relay_module.defined()) << "The IRModule must be defined for the Relay compiler.";
	if (params.size()) {
	ICHECK(relay_module->ContainGlobalVar("main")) << "Missing the main entry function";
	GlobalVar main_glb_var = relay_module->GetGlobalVar("main");
	Function main_func = Downcast<Function>(relay_module->Lookup(main_glb_var));
	auto new_main = BindParamsByName(main_func, params);

This PR moves param binding in VM to OptimizeModule, to be consistent with the graph codegen. Now VMCompiler.optimize(...) returns the graph with constants properly folded.

please review @zhiics @jroesch @icemelon9

[jwfromm]

This makes a lot more sense to me, thanks Masa.

[icemelon]

LGTM.
Could you add a test case for this?

[masahi]

@icemelon9 Added a test to verify that all params are bound in the output of vm.optimize(). This can be checked by counting the number of free vars in the main body. I think this is neat!

[masahi]

@icemelon9 ready to merge