Torch 2.6 is still no good enough for its compiling module

[refraction-ray]

Issue Description

<--! Insert a short description of the bug here, along with what you expected the behavior to be. -->

JIT is still a long way to go in torch, naive torch is much much slower than jax in TC use cases.

How to Reproduce

def vqe(L, d):
    K = tc.backend
    def loss(param):
        c = tc.Circuit(L)
        c.h(range(L))
        for i in range(d):
            c.rzz(range(L-1),range(1, L), theta=param[i, 0][:L-1])
            c.rx(range(L), theta=param[i, 1])
        return K.real(c.expectation_ps(z=[1, 2])+c.expectation_ps(x=[L-2]))
    vgf = K.jit(K.value_and_grad(loss), jit_compile=True)
    param = K.ones([d, 2, L])
    time0 = time.time()
    print(vgf(param)[1].shape)
    time1 = time.time()
    print("jitting time,", time1-time0)
    for _ in range(5):
        vgf(param)
    time2 = time.time()
    print("running time,", (time2-time1)/5)

with tc.runtime_backend("pytorch"):
    vqe(14, 8)

Error Output

InternalTorchDynamoError: AttributeError: 'NoneType' object has no attribute 'make_guard'

Environment Context

<--! Please report your OS version, Python environment and version, TensorCircuit version and necessary dependent package (NumPy, TensorFlow, Jax, Jaxlib, PyTorch) version here. -->

Output of tc.about() and tc.__version__.

OS info: macOS-14.5-arm64-arm-64bit
Python version: 3.12.0
Numpy version: 1.26.4
Scipy version: 1.14.0
Pandas version: 2.2.2
TensorNetwork version: 0.5.0
Cotengra version: 0.6.2
TensorFlow version: 2.17.0
TensorFlow GPU: []
TensorFlow CUDA infos: {'is_cuda_build': False, 'is_rocm_build': False, 'is_tensorrt_build': False}
Jax version: 0.4.31
Jax installation doesn't support GPU
JaxLib version: 0.4.31
PyTorch version: 2.5.0
PyTorch GPU support: False
PyTorch GPUs: []
Cupy is not installed
Qiskit version: 0.46.2
Cirq version: 1.3.0
TensorCircuit version 1.0.0

[refraction-ray]

torch 2.8 passes the above test

[refraction-ray]

a micro benchmark on cpu for three backends, demonstrating the inferior performance of torch backend (though jitting compiling time is much shorter for torch backend, may be a desired feature in some scenarios)

vqe(14, 6)

backend compiling time/running time

tensorflow 19s/0.0017s
jax 48s/0.0013s
pytorch 0.44s/0.048s

[refraction-ray]

remaining issues in pytorch 2.8 backend:

• QR + VMAP
• complex value SVD +AD
• convert_to_tensor + VMAP

[refraction-ray]

also see a new discussion on inferior performance of pytorch for quantum simulation in https://github.com/orgs/tensorcircuit/discussions/116