[python] Use separate context for kernel builder

[lmondada]

This PR fixes the following crash

import cudaq

kernel = cudaq.make_kernel()

@cudaq.kernel()
def kernel2():
    pass

kernel.qalloc(1) # this (or any other MLIR modification) crashes

Error: RuntimeError: the operation has been invalidated

The reason for this is that kernel2 was getting compiled using compile_to_mlir in python/cudaq/kernel/ast_bridge.py. This calls context._clear_live_operations() at the end of the function. However, the context is global, thus it also clears the context for the kernel builder. This invalidates the insertion point stored in the builder, hence any further modifications of the kernel builder fails.

This PR proposes to use a separate context for the kernel builder. When a kernel builder gets invoked and needs to be jitted, we then output the module as string and reload it into the global context. This will come at some performance cost, but only affects kernels defined using the builder and only at the first invocation (as the compiled MLIR module is cached).

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[lmondada]

We'll have to discuss a plan to fix this. For now I've disabled using the kernel builder with kernels that have deferred compilation.