fix: improve cost tracking accuracy for modern models

- Added support for reasoning tokens in cost calculations.
- Fixed DeepSeek cache-write token mapping (PromptCacheMissTokens).
- Improved CalculateCost debug logging to trace all pricing variables.

internal/providers/deepseek.go

@@ -51,6 +51,9 @@ func (u *deepSeekUsage) ToUnified() *models.Usage {
 	if u.PromptCacheHitTokens > 0 {
 		usage.CacheReadTokens = &u.PromptCacheHitTokens
 	}
+	if u.PromptCacheMissTokens > 0 {
+		usage.CacheWriteTokens = &u.PromptCacheMissTokens
+	}
 	if u.CompletionTokensDetails != nil && u.CompletionTokensDetails.ReasoningTokens > 0 {
 		usage.ReasoningTokens = &u.CompletionTokensDetails.ReasoningTokens
 	}