Allow provider-specific model metadata to override canonical values

**Describe the bug**

Recently the `deepseek-ai/deepseek-v4-flash` and `deepseek-ai/deepseek-v4-pro` options were removed from NVIDIA NIM and were replaced with `deepseek-ai/deepseek-coder-6.7b-instruct` and `deepseek-ai/deepseek-v4-flash-0731` and Goose hasn't updated the list yet to reflect these changes.

However, having to wait for a list update in cases like these is very inconvenient and it would be better if we could specify reasoning effort manually.

---

**To Reproduce**

1. Go to 'Switch models'
2. Click first dropdown box and select 'NVIDIA'
3. Click second dropdown box
- The `deepseek-ai` models here are no longer online and the ones that replaced it aren't listed
4. Select 'Enter a model not listed' at the bottom
- You can now input a custom model name, but you can't assign it reasoning effort

---

**Expected behavior**

`deepseek-ai/deepseek-coder-6.7b-instruct` and `deepseek-ai/deepseek-v4-flash-0731` should be in the list.

Custom model should allow the option to specify reasoning effort for times that the list is out-of-date.

---

**Screenshots**

---

**Please provide the following information**

- **OS & Arch:** Windows 10 x64
- **Interface:** GUI
- **Version:** 1.46
- **Provider:** NVIDIA