## Summary
This PR optimizes memory consumption and execution time when assigning a **Chart of Accounts** or **Fiscal Localization**. By moving filtering logic to the database and preventing expensive field prefetching, we've achieved a **60% reduction in peak memory** and cut execution time by more than half on large datasets.

## The Problem
Assigning a chart template was hitting memory limits on databases with a high volume of products (e.g., 2M+). Two main bottlenecks were identified:
* **Inefficient filtering**: Loading all `product.template` <-> `tax` relations into the cache and filtering in-memory using python instead of using SQL.
* **Excessive prefetching**: Accessing `product.product` fields (like `write_date`) inside the compute function triggered a cache miss that prefetched all product fields, consuming significant memory.

## Improvements
* **SQL Filtering:** Pushed the `product_template` filtration logic to the SQL layer to reduce the amount of data loaded into the memory.
* **Prefetching Prevention:** Optimized the compute logic to avoid triggering unnecessary field prefetching on `product.product`.

---

## Benchmarks
*Tested using `memray` on a customer database with ~2 million products.*

| Scenario | Duration | Peak Memory | Total Allocations |
| :--- | :--- | :--- | :--- |
| **Baseline (Before)** | 10:23.4 | 3.6 GB | 9,954,480 |
| **Optimized Prefetching Only** | 10:21.0 | 2.3 GB | 9,292,271 |
| **SQL Filtering Only** | 06:30.2 | 3.0 GB | 8,865,050 |
| **Combined (Final Result)** | **04:59.6** | **1.4 GB** | **8,213,374** |

### Key Results:
* **Memory Saved:** ~2.2 GB (61% reduction)
* **Time Saved:** ~5.5 minutes (52% faster)

OPW-6070666

Forward-Port-Of: odoo/odoo#259304

Introduction of an AI agent to edit the website page being edited.

__What it can currently do__
- Add some snippets to the page and edit them on the fly.
- Make any modifications on the header, footer and wrap element.
- Add some images given by the user or fetched from Unsplash
- Read and write user_custom_rules.scss
- Generate pictures

__What it cannot currently do__
- Change theme and use tools from the editor
- Add some JavaScript to the page
- etc.

__Basic flow__
Every time the user sends a message, the context given in the prompt
is updated with the page information and content as well as the CSS
variable values of the theme.

When asking for a change on the page, the agent can use a tool to
retrieve the HTML of standard snippets, it can also search for images.
When it has everything it needs, it will call the tool to apply changes
on the page. It has a lot of liberty to do so as it can add and replace
any element with arbitrary (sanitized) HTML.

Community PR: odoo/odoo#258133

task-5411100

Purpose:
--------
Allow agents and AI server actions to search for info on the web when answering a user request.

Since Gemini does not support web grounding with tool calls, one added a web grounding tool (and use it for every provider for consistency). This also allows to keep the web search in a separate context with more specific instructions.
The tool will make an independent request to the LLM without any tools but with the web_grounding feature enabled.
LLMs are really bad at recopying URLs and tend to hallucinate links. Therefore, the sources are inserted in the tool result using a uuid, and stored in the session state so that when the main agent references a source, it needs to reference a uuid instead of the source's url. These uuid's are then replaced by the source's urls manually using the mapping stored in the session state.

Web Search can be added on AI agents by using the dedicated Web Search topic, and on server actions of type "AI" by adding the Web Search tool.

The AI APIs request time out was increased to 90 seconds as the web searches might take some time.

### Models restrictions

Some models are now non-accessible by the agents. For example, there's no reason to give the ability to the agent to access private keys; nor to allow it to edit views or access rights. If it is really necessary then these should be dedicated AI tools (eg. to avoid breaking views or to create attachments), or the agent should ask for confirmation to access some values.
There is no real usecase at this moment so the access to these models is currently simply restricted.

By default, every model starting with 'ir.' is restricted. Other models can be restricted by adding them in the AI_MODELS_BLOCKLIST set.

To restrict access to these models, the AI tools **must** use the method `_parse_domain` method defined on `ai.agent`.

Task-6025732