Inconsistency implicit type conversion

**Affects:** Both PythonCall and JuliaCall

**Describe the bug**
I am encountering inconsistent behavior with implicit type conversion when calling a Python function from Julia within a Python file. Specifically, I have observed that the variable 'global_res' is of type 'Julia Float64', whereas res is of type 'Py'. I expected both to be of the same type—Julia Float64—since they are both ultimately handled on the Julia side, where I assumed implicit conversion would occur uniformly. Here is the problematic code segment:
```
from juliacall import Main as jl

jl.seval("""
using PythonCall
""")

jl.global_res = 0

def my_test(x):
jl.global_res = x+1
return x+1

jl.test_pyexec_fn = my_test

jl.seval("""
function jl_add(x::Float64)
res = test_pyexec_fn(x)
println("typeof res:",typeof(res)) #typeof res:Py
println("typeof global_res:",typeof(global_res)) #typeof global_res:Float64
return res
end
""")

x = jl.jl_add(2.0)
```

**Your system**
Information about my system:
- The operating system: Linux
- The versions:
- Julia:1.10.3,
- Python:3.9.19,
- juliacall: 0.9.21,
- PythonCall v0.9.20

Thank you so much in advance!

Start by running the reported Python and Julia reproduction with Julia 1.10.3, Python 3.9.19, juliacall 0.9.21, and PythonCall v0.9.20. Inspect the conversion behavior around test_pyexec_fn, jl.global_res, and jl.jl_add, comparing the reported Py and Float64 results. Done should mean the inconsistency is explained and the intended conversion behavior is fixed or clearly documented.

索引モデルが issue の本文から書いたものです。