[Relay][VM] Relay VM serialization

[zhiics]

This PR implements the serialization and deserialization of the Relay VM. For detailed information, please refer to #3594

An example could be like the following:

import numpy as np
import tvm
from tvm import relay

# define a simple network.
x = relay.var('x', shape=(10, 10))
f = relay.Function([x], x + x)
mod = relay.Module({"main": f})

# create a Relay VM.
ctx = tvm.cpu()
target = "llvm"
compiler = relay.vm.VMCompiler()
vm = compiler.compile(mod, target)
vm.init(ctx)

# serialize.
ser = relay.serializer.Serializer(vm)
code, lib = ser.serialize()

# save and load the code and lib file.
tmp = tvm.contrib.util.tempdir()
path_lib = tmp.relpath("lib.so")
lib.export_library(path_lib)
with open(tmp.relpath("code.bc"), "wb") as fo:
    fo.write(code)
loaded_lib = tvm.module.load(path_lib)
loaded_code = bytearray(open(tmp.relpath("code.bc"), "rb").read())

# deserialize.
deser = relay.deserializer.Deserializer(loaded_code, loaded_lib)
des_vm = deser.deserialize()

# execute the deserialized vm.
des_vm.init(ctx)
x_data = np.random.rand(10, 10).astype('float32')
res = des_vm.run(x_data)
print(res.asnumpy())

@jroesch @wweic @icemelon9 @tqchen @slyubomirsky @MarisaKirisame

[icemelon]

Could we serialize the dst in the beginning since it's the most common field? Also could you add the format of every instruciton in the RFC?

[zhiics]

@icemelon9 yes, we could, but I am not sure if we want to do that because some instructions don't have it. The current order is consistent to the order that they are constructed but the variable-length filed is rotated to the end.

[jroesch]

Can you write a loop over all the testing models load them into vm and serialize them and ensure they produce the same result after deserialization?

[zhiics]

May be just 1 or 2? It may make the testing too slow to have all them in the unit test.

[jroesch]

Can you address my comments otherwise LGTM

[wweic]

lgtm.