[Build][Enhancement] Enable Automatic Kernel Fusion & Integrate FIDESlib GPU Backend into HEIR

## Background: FIDESlib

**FIDESlib** is an open-source, server-side CKKS GPU library designed for high-performance homomorphic encryption on CUDA GPUs, with interoperability to OpenFHE workflows.

- Paper: **"FIDESlib: A Fully-Fledged Open-Source FHE Library for Efficient CKKS on GPUs"**
https://arxiv.org/abs/2507.04775
- GitHub: **CAPS-UMU/FIDESlib**
https://github.com/CAPS-UMU/FIDESlib

---

## Problem Statement

HEIR currently provides mature scheme-level lowering and multiple **library exit paths** (e.g., OpenFHE/Lattigo/Jaxite/tfhe-rs related flows), but there is no GPU backend path for CKKS that directly targets a CUDA library such as FIDESlib.

Because of this gap:

1. It is hard to lower CKKS programs from HEIR into a GPU-native execution path.
2. There is no dedicated IR stage to make kernel-fusion decisions before backend codegen.
3. End-to-end fused kernel emission against a CKKS GPU library is not available yet.

---

## Why This Contribution

Based on the above, I would like to contribute a project to HEIR that adds:

1. A fusion-friendly intermediate representation for CKKS GPU lowering.
2. Conversion passes into a FIDESlib-oriented exit representation.
3. A kernel-fusion pass that decides what to fuse.
4. Code generation that can emit CUDA code (including fused kernels when needed) for FIDESlib execution.

---

## Summary of Proposed Extension

I propose a GPU backend extension path for HEIR targeting CKKS workloads with FIDESlib integration:

1. New `limb` dialect + `lwe-to-limb` conversion
2. A kernel fusion pass on `limb` IR (horizontal + vertical fusion)
3. New `fideslib` dialect + `limb-to-fideslib` conversion
4. `heir-translate` extension for `.cuh` / `.cu` emission and fused kernel codegen

---

## Proposed Design

### 1) `limb` dialect + `lwe-to-limb` conversion pass

#### Purpose
Bridge existing LWE-level representation to a fusion-friendly, low-level CKKS representation.

#### Dialect content (initial)
- Primitive ops (initial scope):
- modular add/sub
- modular mul/reduce
- NTT / INTT
- rotate
- Types/attributes needed for fusion planning
- `fusion_id` (or equivalent) attribute for downstream fused-kernel grouping

#### Pass
- `lwe-to-limb` conversion to map existing LWE ops/types into `limb` equivalents.

---

### 2) Fusion pass on `limb` dialect

#### Scope
Implement transform logic for:
- Horizontal fusion
- Vertical fusion

#### Initial algorithm
1. Build graph over `limb` ops
2. Assign / track memory buffers per node
3. Identify candidates sharing/reusing compatible buffers
4. Score candidates using expected memory savings
5. Apply highest-benefit legal fusion(s)

---

### 3) `fideslib` dialect + `limb-to-fideslib` conversion pass

#### Purpose
Provide an exit dialect representing FIDESlib-facing operations and API-level semantics.

#### Dialect content (initial)
- Ops mapped to FIDESlib-supported operations
- Types for FIDESlib context/config and API arguments
- `fusion_id` propagation for fused codegen

#### Pass
- `limb-to-fideslib` conversion to lower fusion-ready `limb` IR into FIDESlib-level IR.

---

### 4) Codegen (`heir-translate`)

#### Codegen
Extend `heir-translate` to emit:
- header files (`.cuh`)
- source files (`.cu`)

Behavior:
- Emit direct FIDESlib calls for supported kernels
- Emit generated fused kernels when fusion patterns are not natively provided by FIDESlib