Split Generated Code into Multiple Files

## :rocket: Feature Request

### Affected Languages

- [ ] `TypeScript` or `Javascript`
- [x] `Python`
- [ ] `Java` (haven't checked)
- [ ] .NET (`C#`, `F#`, ...) (haven't checked)

### General Information
* **JSII Version:** 1.17

* [x] I may be able to implement this feature request

### Description

The `cdktf` provider for AWS generates a 19 MB `__init__.py` file which contains all the classes for the Terraform AWS Provider. You can check it with this [project](https://github.com/terraform-cdk-providers/cdktf-provider-aws) which publishes a prebuilt PyPi [package](https://pypi.org/project/cdktf-cdktf-provider-aws/).

According to this [issue](https://github.com/hashicorp/terraform-cdk/issues/504), this might lead to a subpar development experience in certain IDEs. At least in [PyCharm](https://www.jetbrains.com/pycharm/), it doesn't work out of the box. After setting custom indexing settings for intellisense it seems to work, but the indexing process is still slow. I don't know how other IDEs are behaving.

Copied from the [issue](https://github.com/hashicorp/terraform-cdk/issues/504):

> Change the system property idea.max.intellisense.filesize (https://www.jetbrains.com/help/pycharm/tuning-the-ide.html#configure-platform-properties) to enable intellisense for the imports.aws module, but notice how slow the indexing operation is, because of the large file size.

### Proposed Solution

Without having looked at this in detail, one option could be to somehow provide an option to create submodules for each input class / namespace. Is this possible already, can't really tell from the [docs](https://github.com/aws/jsii/blob/377eae2058bc09e5350a17ab5e8fac627e1fe688/gh-pages/content/specification/2-type-system.md#submodules)?

This is possibly related to https://github.com/aws/jsii/issues/1919

Start by tracing how the Python cdktf provider produces its generated __init__.py file, using the linked provider project as the reproduction case. Compare the current single-file output with the proposed submodule-per-input-class or namespace approach, and verify that the resulting package improves IDE indexing without breaking imports.

Written by the indexing model from the issue text.